数据集首页 · 魔搭社区

systemd — Tue, 20 Oct 2026 21:21:08 GMT

The structures behind biomolecular design.
Indexed molecular records, sampling metadata, and traceable source inventories.

Model · Inventory · Training guide

RFD4-Proteína training datasets

Experimental (PDB) and predicted (AlphaFold DB) molecular structures packaged for RFD4-Proteína training: AFDB protein monomers and PDB interfaces, including protein–DNA, antibody and TCR interfaces. Original structure shards, byte offsets, indices, and sampling tables are preserved.

This release contains only PDB-derived and AlphaFold DB-derived components. Other components of the full training mix (distilled and synthetic sets, small-molecule, RNA and protein–ligand collections) are not part of this release, and neither are components that were derived from them. Dataset availability here does not establish which components a particular checkpoint used.

Looking to generate designs? Start with the model card. You do not need this dataset to run the model.

Snapshot at a glance

	Details
Configurations	5 named metadata views over 2 blob directories (AFDB monomers, PDB snapshot)
Payload	4,251 objects · 238.52 GB (222.15 GiB) across the three published component directories
Storage	Parquet sampling tables and indices, plus independently compressed records in binary shards
PDB snapshot	October 4, 2026 PDB snapshot, annotated with CATH and PLIP
License	CC BY 4.0 for the added annotations, sampling tables and indices; upstream terms apply to the structures (see Source terms and attribution)

The source snapshot date is not a universal model-training cutoff.

Choose a component

Counts and proposed sampling shares come from dataset-inventory.json. “Passing filters” refers to the recorded audit filters, not an independently revalidated quality score.

Configuration	Contents	Published examples	Passing filters	Proposed share in the full mix
`afdb_monomer`	AFDB protein monomers	1,186,995	914,831	69%
`pdb_interfaces`	PDB molecular interfaces	13,938,465	8,431,727	14.5%
`antibody_interfaces_pdb`	PDB antibody interfaces	81,895	72,562	0.5%
`tcr_interfaces_pdb`	PDB TCR interfaces	9,526	8,608	0.5%
`pdb_dna_interface`	PDB protein–DNA interfaces	12,812	—	—

The shares are preserved from the source manifest of the full training mix and are not renormalized for this subset; they sum to less than 100% here because the other components are not released. pdb_dna_interface has no recorded share. These are overlapping sampling views, not counts of globally unique structures. The antibody/TCR PDB views share immune_interfaces.parquet and are distinguished by runtime filters; the general PDB view also shares their blob directory, and pdb_dna_interface reads structures from the same PDB blobs.

Load metadata first

For inspection without downloading structure shards:

pip install datasets huggingface_hub

from datasets import load_dataset

rows = load_dataset("nvidia/RFD4-Proteina", "afdb_monomer", split="train", streaming=True)
print(next(iter(rows)))

The rows contain metadata, not decoded molecular structures. load_dataset does not apply the manifest's runtime filters, cluster weighting, or proposed sampling shares. The train label describes the exposed sampling table; it does not imply a leakage-audited train/test split for your task.

Download and read molecular records

The format separates which example to sample from where its structure is stored:

component/
├── examples.parquet or metadata.parquet   # sampling and quality metadata
└── blob/
    ├── index.parquet                     # record ID → shard, offset, length
    └── data/*.bin                        # independently compressed records

Shards are at most about 67 MB, so a single record can be read without downloading a component. From an environment with the RFD4/AtomWorks dependencies installed:

from pathlib import Path
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download
from atomworks.ml.utils.blob_store import BlobStore

repo = "nvidia/RFD4-Proteina"
prefix = "train_datasets/streaming-latest/synthetic/afdb_genie3_subset/"
index = hf_hub_download(repo, prefix + "blob/index.parquet", repo_type="dataset")
row = pq.read_table(index).slice(0, 1).to_pylist()[0]
shard = hf_hub_download(repo, prefix + "blob/data/" + row["shard"], repo_type="dataset")
store = BlobStore(str(Path(shard).parent))
record = store.get_bytes(row["shard"], int(row["offset"]), int(row["length"]))
Path("example.cif").write_bytes(record)
print(row["example_id"], "saved to example.cif")

This reads one indexed structure, not a training-sampler draw. BlobStore handles per-record decompression. For training, use the component's matching sampling table and the loader/configuration in the training guide.

AtomWorks' indexed-blob loader supports local storage and S3 ranged reads. Direct Hugging Face ranged reads are not a drop-in loader backend; download the selected component locally. Keep its index and shards from the same revision.

Filters, provenance, and reproducibility

The inventory records each component's sampling table, blob directory, filters, proposed share, and object sizes. Component READMEs and provenance/ folders hold additional processing details.

For a reproducible training run, record the dataset commit, model/code revisions, component selection, runtime filters, mixture weights, and sampler settings. File names containing latest do not pin data; the Hub commit does. Do not mix metadata from one revision with shards from another.

Important component qualifications

PDB interfaces: recorded filters include deposition before January 1, 2024, resolution below 9 Å, and the inventory's bond, ligand, interface, and size gates. Inherited interface definitions and geometry-dependent flags were not recertified on new coordinates.
PDB–DNA interfaces: added on 2026-10-13; the table (12,812 rows) reads structures from the existing PDB blobs, with matching dna_binding_interfaces_20261013 sampling files.
AFDB monomers: predicted structures filtered to at most 256 standard amino-acid residues; they are not experimental measurements.
Missing counts: null means unknown or unaudited, not zero. Read component metadata for the precise atom/residue-count conventions before filtering.

Intended use and limitations

Use this snapshot for training, fine-tuning, data inspection, and reproducibility studies with compatible molecular loaders. Predicted structures are not experimental measurements. Coverage and errors inherit the biases of the upstream experimental and predicted sources. Large numbers of rows do not imply equivalent diversity or independent samples.

Before constructing an evaluation, audit overlap with your targets at the appropriate sequence, structure, complex, and source-date levels. Neither the Hub metadata configuration nor the proposed mixture automatically performs that audit.

Benchmarks and project media

Besides the training data, this repository hosts files that the code repository downloads instead of storing:

benchmarks/ae_pdb_temporal_holdout_2026_10_04/: the autoencoder validation set. 640 annotated PDB structures (128 each of protein–ligand, modified residue, glycan, DNA and RNA examples, all deposited on or after 2024-01-01) as per-file Zstandard-compressed mmCIF, with a README describing the selection.
benchmarks/nanobody/: 160 nanobody design validation inputs, 40 target structures crossed with 4 VHH frameworks (caplacizumab, ozoralizumab, vobarilizumab, h-NbBCII10).
metrics/cdr_torsion_data/: reference data for the cdr_torsion_naturalness antibody metric: native CDR phi/psi samples (VH/VL and VHH) and their kernel-density parameters. The metric downloads it on first use.
media/: images and video used by the project page and the code repository's README. This is not training data.

make setup in the code repository downloads the two benchmark directories into data/benchmarks/. They are not part of the load_dataset configurations above.

Source terms and attribution

The structures keep the terms of their upstream sources:

PDB entries come from the wwPDB archive, whose data are available under the CC0 1.0 waiver.
AlphaFold DB structures (EMBL-EBI / Google DeepMind) are available under CC BY 4.0, which requires attribution.
Derived annotations such as CATH domain labels (CC BY 4.0) and PLIP interaction annotations come from their own sources; cite them when you use them.

The sampling tables, indices, annotations and packaging added for RFD4-Proteína are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0). All upstream sources above are compatible with CC BY 4.0 attribution. The model's BSD-3-Clause license is separate and does not govern these data. Credit the upstream sources as well as this dataset, and check the component README before reuse.

When reporting work, cite the RFD4-Proteína dataset repository at the exact commit, credit each upstream source used, and describe any filtering or resampling. The inventory and component READMEs describe how each component was processed.

Model and examples · Dataset history · Code and issues

systemd — Tue, 20 Oct 2026 21:19:20 GMT

数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。

当前数据集卡片使用的是默认模版，数据集的贡献者未提供更加详细的数据集介绍，但是您可以通过如下GIT Clone命令，或者ModelScope SDK来下载数据集

下载方法

:modelscope-code[]{type="sdk"} :modelscope-code[]{type="git"}

systemd — Tue, 20 Oct 2026 21:17:49 GMT

Nemotron-Personas-UK

A compound AI approach to personas grounded in real-world distributions

Dataset Overview

Nemotron-Personas-UK is an open-source (CC BY 4.0) dataset of synthetically-generated personas grounded in real-world demographic, geographic, and personality trait distributions in the United Kingdom, to capture the diversity and richness of the population. This dataset was built in collaboration with Locai Labs, the sovereign AI computer company based in London. It's the first UK dataset of its kind aligned with statistics for names, sex, age, ethnicity, religion, marital status, household type, education, employment, occupation, industry, pay, health, language (including Welsh, Gaelic, Scots, Irish and Ulster-Scots), carer status and location. Not all of these are published as fields; see What is NOT in the dataset. The dataset covers all four nations. England and Wales are grounded in Census 2021 from the Office for National Statistics (ONS), Scotland in Scotland's Census 2022 from National Records of Scotland (NRS), and Northern Ireland in Census 2021 from the Northern Ireland Statistics and Research Agency (NISRA). Each census has its own probabilistic graphical model (PGM), and the three are combined into one dataset.

Nemotron-Personas-UK supports UK model builders in developing Sovereign AI systems that incorporate important region-specific demographics, language, and cultural context. The dataset improves diversity of synthetic data generation, mitigates data and model biases, and helps prevent model collapse (degradation caused by uncurated training on another model's outputs) by reflecting the UK's real geographic and demographic distributions across England and Wales' 329 local authority districts, Scotland's 32 council areas, and Northern Ireland's 11 local government districts. In particular, the dataset is designed to be more representative of underlying demographic distributions along multiple axes, including regional diversity (e.g., major cities, rural areas, and the four nations), ethnicity, religious affiliation, household composition, education levels, and a wide array of occupations, as compared to past persona datasets. As an example, one can produce high-quality multi-turn chat conversation data with realistic names, ages, and backgrounds that reflect the authentic experience of people across the UK, bringing unique perspectives and regional nuances to that data.

Produced using NeMo Data Designer, an enterprise-grade compound AI system for synthetic data generation, the dataset leverages a Probabilistic Graphical Model (PGM) for each census, along with the Apache-2.0-licensed google/gemma-4-31B-it model and an ever-expanding set of validators and evaluators built into Data Designer. The dataset was generated using 18 NVIDIA GPUs, hosted across a hybrid set-up of cloud and local compute, using Locai Labs' own AI computer, the Locai 1. An extended version of Nemotron-Personas-UK will be available for use in NeMo Data Designer itself.

This dataset is ready for commercial or non-commercial uses.

What is NOT in the dataset

Given the emphasis on personas, the dataset excludes other fields available in NeMo Data Designer, e.g., first and last names, detailed personas, Big Five personality traits and synthetic addresses, as well as more sensitive fields such as ethnicity, religion, country of birth and disability. Also excluded are personas generally of relevance to enterprise clients, such as finance and healthcare. Please reach out to explore enterprise use cases.

All data, while mirroring real-world distributions, is completely artificially generated. Any similarity in names or persona descriptions to actual persons, living or dead, is purely coincidental.

Data Developer

Locai Labs & NVIDIA Corporation

Release Date

Released on Hugging Face on October 20, 2026

https://huggingface.co/datasets/nvidia/Nemotron-Personas-UK

Dataset Creation Date

October 12, 2026

License/Terms of Use

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

The personas are synthetic, built from the public sources below under each source's license.

Under the Open Government Licence v3.0:

Source: Office for National Statistics licensed under the Open Government Licence v.3.0 (Census 2021; Annual Survey of Hours and Earnings; Annual Population Survey and Labour Force Survey; Rural Urban Classification 2021; Index of Place Names, July 2024; baby names).
© Crown copyright. Data supplied by National Records of Scotland (Scotland's Census 2022; name statistics).
Source: NISRA: Website: www.nisra.gov.uk (Census 2021; Annual Survey of Hours and Earnings 2024; name statistics).
Contains OS data © Crown copyright and database right 2024.
Contains Royal Mail data © Royal Mail copyright and database right 2024.

Under CC BY 4.0:

Name statistics from the Central Statistics Office, Ireland.
Big Five trait means: Rentfrow, P. J., Jokela, M., and Lamb, M. E. (2015). Regional personality differences in Great Britain. PLOS ONE 10(3): e0122245. doi:10.1371/journal.pone.0122245
Names from other Nemotron-Personas datasets (NVIDIA).

Other sources:

Contains information from Companies House: People with Significant Control data, aggregated name counts only.
US first names from the Social Security Administration (public domain).

Use Case

Developers working on Sovereign AI, training LLMs, seeding user simulations to test and evaluate AI agents, and/or looking to improve diversity of synthetically generated data, mitigate data/model biases, and prevent model collapse.

Data Version

1.0 (October 20, 2026)

Intended Use

The Nemotron-Personas-UK dataset is intended to be used by the community to continue to improve open models and push the state of the art. The data may be freely used to train any model. We welcome feedback from the open-source community and invite developers, researchers, and data enthusiasts to explore the dataset and build upon it.

The Nemotron-Personas-UK dataset is grounded in distributions of self-reported demographic data in Census 2021 for England and Wales, Scotland's Census 2022, and Census 2021 for Northern Ireland. As such, its primary goal is to support Sovereign AI development by combating missing data and/or potential biases present in model training data today, especially when it comes to existing persona datasets used in synthetic data generation. Despite the improved data diversity and fidelity to the UK's population, we are still limited by data availability and reasonable model complexity. This results in some necessary independence assumptions. The three censuses use different questions and categories, and some, such as qualifications, do not map exactly between nations. These categories were harmonised through mapping to a coarser set of categories for each of the relevant fields; marital status was mapped to a unified 5-category scheme, household type to 13 categories, and qualifications to 6. Published census tables also suppress small counts to protect privacy, which prevents the tail of the distribution of census data from being represented. We mitigate this issue through Iterative Proportional Fitting (IPF) of the fine-grained individual census tables to match the broader cross tables which are otherwise suppressed. Sensitive attributes such as ethnicity, religion and disability status are synthetic, sampled from published census distributions, and describe no real person. They are used to condition the persona text but are not published as fields in this release. They must not be used to infer, profile, or make decisions about real individuals or groups. Where age-related signals are derived from public records (for example, Companies House nationality name statistics), only decade-level aggregates are used; no individual dates of birth are retained. We leave further efforts to improve fidelity to future work.

Note that the dataset is focused on adults only (18+ years).

Dataset Details

The dataset contains:

6M personas in English across 1M records (889k records for England and Wales, 83k for Scotland and 28k for Northern Ireland)
23 fields: 6 persona fields, 6 persona-attribute fields, 10 contextual fields grounded in official demographic and labour statistics, and a unique identifier
618M tokens total, including 242M persona tokens
4 nations, 12 regions and 372 local authorities
467k unique full names (27k unique first names and 25k unique last names)
104 occupation groups (SOC 2020 minor groups) across 86 industry categories (SIC 2007 divisions), all grounded in real-world distributions
Comprehensive coverage across demographic, geographic, and personality trait axes
A variety of persona types: professional, sports, arts, travel and culinary, plus a concise general-purpose persona
Natural language persona attributes: cultural background, skills & expertise, goals & ambitions, hobbies & interests

Nemotron-Personas-UK was designed to align with the UK's official demographic and labour statistics (ONS Census 2021, Scotland's Census 2022, and NISRA Census 2021), while extending them into areas important for AI training. In practice, this meant:

Granular Geographic Distribution: the dataset anchors records in 372 local authorities across England, Wales, Scotland, and Northern Ireland, mirroring the real population density and distribution across the UK.
Naming Conventions: by leveraging first-name and surname statistics from ONS, NRS, and NISRA, plus Companies House People with Significant Control (PSC) first names and surnames by nationality, the Central Statistics Office of Ireland, US Social Security Administration names, and names from other Nemotron-Personas datasets, all of which were classified using gemma-4-31B-it by which country the name is likely to be from, the dataset incorporates 27k unique first names and 25k unique last names, reflecting the linguistic and ethnic diversity of the UK population.
Standardised Occupational Framework: professional data is aligned with the ONS Standard Occupational Classification 2020 (SOC 2020) and the UK Standard Industrial Classification 2007 (SIC 2007), ensuring that the workforce distribution in the dataset matches official labour market statistics. Weekly pay is drawn from the ONS Annual Survey of Hours and Earnings (ASHE) and Annual Population Survey / Labour Force Survey tables, as well as NISRA ASHE 2024 tables.
Demographic Alignment: key variables such as age, sex, ethnicity, religion, marital status, household type, education level, economic activity, occupation, industry and long-term health and disability are sampled to reflect the current demographic profile of each nation, providing a statistically representative synthetic population.
Multi-Census Integration: the England and Wales, Scotland, and Northern Ireland PGMs each follow their own census schema, with counts in proportion to the real adult population share of each nation within the UK (84.3% England, 8.32% Scotland, 4.63% Wales, and 2.75% Northern Ireland).
Personality Traits: Big Five (OCEAN) T-scores by local area are drawn from Rentfrow, Jokela and Lamb (2015), in particular the S2 table. Northern Ireland, which was excluded from the paper, uses the average of the other nations' T-scores. The traits condition the persona text but are not published as fields.
Sampling: all 1M generated records are released. Regeneration of the prose was performed on records found to have issues through manual and automatic review.

These extensions build on a strong foundation of public data, helping create personas that are both statistically grounded and culturally representative, while remaining synthetic, privacy-preserving, and open.

Seed Data

In order to capture the socio-demographic and geographic diversity and complexity of the UK's population, Nemotron-Personas-UK leveraged:

Census 2021 for England and Wales, published by the Office for National Statistics (ONS)
Scotland's Census 2022, published by National Records of Scotland (NRS)
Census 2021 for Northern Ireland, published by the Northern Ireland Statistics and Research Agency (NISRA)
All census data from the sources above is from their respective table builders (ONS, Scotland's Census, NISRA), which allow a user to download any cross table which is not subject to confidentiality controls
Earnings data from the ONS Annual Survey of Hours and Earnings (ASHE) and published Annual Population Survey / Labour Force Survey tables
First and last name distributions from ONS, NRS, and NISRA name statistics, plus Companies House People with Significant Control (PSC) first names and surnames by nationality, the Central Statistics Office of Ireland, US Social Security Administration names, and names from other Nemotron-Personas datasets
Big Five personality trait means by local area from Rentfrow, Jokela and Lamb (2015), PLOS ONE
ONS Census 2021 custom cross-tabulation tables (English proficiency by industry/occupation; religion and main language by country of birth), England and Wales
NISRA Census 2021 Flexible Table Builder, Northern Ireland (English proficiency by industry/occupation)
ONS Rural Urban Classification 2021 (holiday-destination weighting)
ONS Index of Place Names in Great Britain, July 2024 (place-name validation in travel text)
ONS baby names, Wales tables 1997 to 2005 (Welsh first names)
Companies House People with Significant Control (Welsh-origin first names and surname consistency; counts only, aggregated)

Schema

The dataset includes 23 fields: 6 persona fields, 6 persona-attribute fields, 10 contextual fields, and a unique identifier, as shown below. Researchers will find many contextual fields useful in zeroing in on specific demographics, which is challenging to do with existing datasets. Categorical values use harmonised UK-wide labels rather than each census' published labels where the field categories differ between countries; otherwise they follow the published labels.

Nemotron-Personas-UK
  |-- uuid: string                        Globally unique identifier
  |-- professional_persona: string        Professional persona: field of work, key skills and traits
  |-- sports_persona: string              Sports persona: athletic interests, teams, approach to fitness
  |-- arts_persona: string                Arts persona: engagement with creative expression and the arts
  |-- travel_persona: string              Travel persona: travel interests and style
  |-- culinary_persona: string            Culinary persona: food preferences, cooking and dining habits
  |-- persona: string                     Concise general-purpose persona capturing the person's essence
  |-- cultural_background: string         Description of the person's cultural background
  |-- skills_and_expertise: string        Professional and personal skills in narrative format
  |-- skills_and_expertise_list: string   List of skills and areas of expertise (JSON-encoded list)
  |-- hobbies_and_interests: string       Interests and recreational activities in narrative format
  |-- hobbies_and_interests_list: string  List of hobbies and personal interests (JSON-encoded list)
  |-- career_goals_and_ambitions: string  Professional aspirations and long-term career objectives
  |-- sex: string                         Sex (Male, Female)
  |-- age: integer                        Age in years (18+)
  |-- marital_status: string              Legal partnership status (5 harmonised categories)
  |-- household_type: string              Type of household the person lives in (13 harmonised categories)
  |-- education_level: string             Highest level of qualification (6 UK-wide levels)
  |-- occupation: string                  SOC 2020 minor group (code and title) or Does not apply
  |-- industry: string                    SIC 2007 division (code and title) or Does not apply
  |-- municipality: string                Local authority of residence: local authority district (England and Wales),
  |                                       council area (Scotland) or local government district (Northern Ireland)
  |-- region: string                      One of 12: the 9 English regions, plus Wales, Scotland and Northern Ireland
  |-- country: string                     Constant: United Kingdom

Field & Token Counts

The dataset consists of 618M tokens (242M persona tokens) across 1,000,000 records and 22 columns, excluding the globally unique identifier, counted with the GPT-3.5-turbo tokenizer (tiktoken cl100k_base). Note that the data provides comprehensive coverage across England, Wales, Scotland, and Northern Ireland and 372 local authorities (329 local authority districts in England and Wales, 32 council areas in Scotland and 11 local government districts in Northern Ireland). The table below lists, for each column, the number of unique values, missing values, the data type, and average and total tokens.

column	num_unique	type	avg_tokens	total_tokens
`uuid`	1,000,000	string	N/A	N/A
`professional_persona`	1,000,000	string	44.256	44,255,556
`sports_persona`	1,000,000	string	38.428	38,428,235
`arts_persona`	999,998	string	42.025	42,025,218
`travel_persona`	1,000,000	string	42.171	42,170,572
`culinary_persona`	999,996	string	42.877	42,877,183
`persona`	1,000,000	string	32.029	32,028,540
`cultural_background`	1,000,000	string	73.166	73,166,077
`skills_and_expertise`	1,000,000	string	68.800	68,800,080
`skills_and_expertise_list`	981,660	string	29.404	29,404,321
`hobbies_and_interests`	1,000,000	string	76.606	76,606,349
`hobbies_and_interests_list`	999,413	string	32.169	32,169,193
`career_goals_and_ambitions`	999,996	string	60.086	60,085,832
`sex`	2	string	1.000	1,000,000
`age`	82	int64	1.000	1,000,000
`marital_status`	5	string	3.878	3,878,310
`household_type`	13	string	8.359	8,358,736
`education_level`	6	string	4.186	4,186,300
`occupation`	105	string	4.565	4,564,669
`industry`	87	string	5.823	5,822,823
`municipality`	372	string	3.224	3,223,963
`region`	12	string	2.207	2,207,388
`country`	1	string	2.000	2,000,000

Dataset Description & Quality Assessment

The analysis below provides a breakdown across various axes of the dataset to emphasize the built-in diversity and pattern complexity of the UK demographic data.

Names

Since the focus of this dataset is on personas, names aren't provided as dedicated fields. However, infused into persona generation are 27,100 unique first names and 24,929 unique last names, forming 466,865 unique full names, grounded in ONS, NRS, and NISRA name statistics, Companies House People with Significant Control (PSC) first names and surnames by nationality, the Central Statistics Office of Ireland, US Social Security Administration names, and names from other Nemotron-Personas datasets. First names are drawn by sex, decade of birth and country of birth; people born abroad or of non-British heritage draw from their birth country's names, filtered to fit their stated religion. Surnames are drawn by country of birth or heritage and conditioned on the first name using Companies House first-name by surname pairs. Every name is held by at least 10 Companies House holders. Names are a proxy, not a census.

Welsh-born Welsh speakers can receive Welsh-origin names. Rare names suppressed in published statistics are not restored; supplementary sources cover them.

Age Distribution

The age distribution of Nemotron-Personas-UK mirrors the country's real demographic structure. The median age is 48. The adult population is spread evenly from 25 to 64, at 16–17% per ten-year band; 18–24 makes up 10.6%, 65–74 12.6% and 85+ 3.0%. Women are 51.7% overall, rising to 54.9% at 75–84 and 62.8% at 85+.

Note that minors under 18 are excluded from this dataset.

Marital Status by Age Group

The heatmap below displays the fraction of people for each age cohort by marital or civil partnership status, out of Never married, Married or civil partnership, Separated, Divorced or dissolved, and Widowed or surviving partner. Marriage becomes the most common status at 35–44 (53.6%) and peaks at 65–74 (64.4%); divorce peaks at 55–64 (17.0%); widowhood dominates from 85 (59.3%). All of these considerations are relevant to informing authentic life experiences and personas in the UK context.

Household Type by Age Group

The heatmap below shows, for each age cohort, the type of household personas live in (note that this describes the household, not the persona's own role in it). 18–34-year-olds mostly live in 'other' households (sharers or adults living with parents) or as couples; married couples with dependent children are most common at 35–54 (40.9% at 35–44); from 55, couples whose children have left home dominate; from 65, pensioner families; at 85+, half live alone (50.2%) and 11.3% live in communal establishments.

Education Level by Age Group

The heatmap below captures patterns of educational attainment (highest level of qualification) across age cohorts in the UK. Degree-level qualifications peak at 25–34 (49.0%) and fall with age; 'no qualifications' rises from 8.5% at 25–34 to 54.4% at 85+. Each census' qualifications are mapped to six UK-wide levels; Scottish qualifications use the nearest England and Wales level, with HNC/HND counted as degree level.

Geographic Intricacies of Education Attainment

This slice of the dataset demonstrates how geography informs education and therefore persona descriptions. The choropleth map shows, for each municipality, the share of residents aged 25 and older who hold a degree-level qualification, relative to the national average of 37.1%. The highest areas are Richmond upon Thames (65.9%), Wandsworth (65.4%) and Camden (64.5%), and the lowest are Great Yarmouth (19.0%), Fenland (19.6%) and Boston (20.0%). By region, London is highest (49.3%) and the North East lowest (29.9%).

Occupational Categories

The bar chart below reflects the professional distribution of the personas in work (612,488, 61.2% of all records) across major occupation groups. These groups are grounded in the nine major groups of SOC 2020 and are further informed by demographic and geographic distributions. The largest groups are Professional occupations at 21.1%, Associate professional and technical occupations at 12.8%, and Managers, directors and senior officials at 11.6%. 87% of professionals hold a degree-level qualification, against 16% in elementary occupations and 15% of process, plant and machine operatives, where approximately 25% have no qualifications.

Persona Diversity

How to use it

You can load the dataset with the following lines of code.

from datasets import load_dataset

nemotron_personas = load_dataset("nvidia/Nemotron-Personas-UK", split="train")

Dataset Characterization

Data Collection Method

Hybrid: Human, Synthetic, Automated

Labeling Method

Not Applicable

Dataset Format

All fields are strings except age (integer); skills_and_expertise_list and hobbies_and_interests_list are JSON-encoded lists. The full data structure is provided in the Schema section.

Dataset Quantification

Record counts: 1,000,000 records (6,000,000 personas)
Total data storage: 1.5 GB (7 Parquet shards)

Limitations

In Northern Ireland and west-central Scotland, sports personas may reflect the general association between some football clubs and community background. No data source of club support by religion or community is used, and no claim is made about any individual.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Citation

If you find the data useful, please cite:

@software{nvidia/Nemotron-Personas-UK,
  author = {Stephens, Samuel and Drayson, George and Herr, Nathan and Drayson, James and Drayson, Paul and {Kabiri Chimeh}, Mozhgan and Sharabiani, Ashton and Sadeghi, Bardiya and Jennings, Will and Meyer, Yev},
  title = {{Nemotron-Personas-UK}: Synthetic Personas Aligned to Real-World Distributions for the United Kingdom},
  month = {October},
  year = {2026},
  url = {https://huggingface.co/datasets/nvidia/Nemotron-Personas-UK}
}

systemd — Tue, 20 Oct 2026 21:17:10 GMT

OSWorld judge probe

Can a model tell whether a computer-use agent really finished an OSWorld task?

This is not an OSWorld task-success score. Each model reads a finished Claude Opus 5 episode as text (the agent's narration and actions, no screenshots) and predicts whether the official OSWorld checker gave it full credit. Labels are the checker's published rewards.

OSWorld Benchmark v1 (OSWorld-Verified, n = 358, task instruction shown)

Model	AUROC (95% CI)	Brier ↓	Acc @0.5	Mean P(complete)
Drex Max v1	0.780 (0.713–0.841)	0.114	86.0%	78.2%
Qwen3.8-27B	0.770 (0.693–0.841)	0.112	86.0%	80.8%
GPT-6 Sol	0.744 (0.672–0.815)	0.114	86.3%	86.8%
Jev 1.13.0	0.734 (0.657–0.804)	0.240	61.2%	49.2%
GPT-6 Luna	0.709 (0.631–0.782)	0.148	83.0%	81.8%
Drex v2.0	0.698 (0.624–0.772)	0.133	81.8%	77.0%
Shorter-transcript heuristic	0.709	0.140 (base rate)	—	83.2% (true)

OSWorld Benchmark v2 (OSWorld 2.0, n = 105, no instruction — V2 tasks are gated)

Model	AUROC (95% CI)	Brier ↓	Acc @0.5	Mean P(complete)
GPT-6 Sol	0.750 (0.644–0.851)	0.259	59.0%	51.0%
Drex Max v1	0.737 (0.629–0.836)	0.189	68.6%	40.5%
Qwen3.8-27B	0.687 (0.576–0.793)	0.287	49.5%	58.7%
Jev 1.13.0	0.687 (0.564–0.803)	0.188	73.3%	31.2%
GPT-6 Luna	0.653 (0.534–0.769)	0.290	60.0%	47.8%
Drex v2.0	0.636 (0.514–0.753)	0.409	36.2%	73.9%
Shorter-transcript heuristic	0.694	0.208 (base rate)	—	29.5% (true)

AUROC is computed on items answered by every model. Confidence intervals are 2,000 bootstrap resamples (seed 0).

Data

Source episodes: xlangai/ubuntu_osworld_verified_trajs
- v1: results_opus5_task221874_osworld1.0_shareable.zip (358 scored; instructions from xlang-ai/OSWorld evaluation_examples)
- v2: results_opus5_task220297_osworld2.0_shareable.zip (105 scored)
data/osworld_{v1,v2}_outcome.jsonl: System One requests with two questions:
- complete (noul): did the checker give full credit?
- outcome (choice): failed / partial / complete.
Transcripts over 80k characters keep the first 20k and last 60k.

Predictions

predictions/<set>_<engine>.jsonl: one row per episode, p = option probabilities (noul [false, true]; choice in criteria order).

drexmax: local weights nace-ai/drex-max-v1 (probe_drexmax.py)
qwen38-27b: Fireworks dedicated deployment of qwen3p8-27b, JSON-prompted (probe_llm.py)
gpt6sol: OpenAI gpt-6-sol, JSON-prompted (probe_llm.py)
jev: jev-1.13.0 via /v1/systemone (probe_api.py)
gpt6luna: OpenAI gpt-6-luna, JSON-prompted (probe_llm.py)
drexv2: drex-v2.0 via /v1/systemone (probe_api.py)

Reproduce

python code/build_probe.py v1 && python code/build_probe.py v2
# score engines (see each script's docstring), then:
python code/score.py v1 && python code/score.py v2
python code/plot_auroc.py

systemd — Tue, 20 Oct 2026 21:17:09 GMT

Nemotron-Personas-Germany

Ein Compound-AI-Ansatz für synthetische Personas auf Basis realer Verteilungen
A compound AI approach to personas grounded in real-world distributions

Datensatzübersicht (Dataset Overview)

Nemotron-Personas-Germany ist ein Open-Source-Datensatz (CC BY 4.0) mit synthetisch generierten Personas, die auf realen demografischen und geografischen Verteilungen in Deutschland basieren, um die Vielfalt der Bevölkerung abzubilden. Es ist der erste deutsche Datensatz dieser Art, der auf Statistiken zu Namen, Geschlecht, Alter, Migrationshintergrund, Familienstand, Haushaltsform, Bildung, Erwerbstätigkeit, Beruf und Wohnort basiert. Die Daten wurden unter Verwendung offizieller Statistiken aus der deutschen Volkszählung 2022 (Zensus 2022) synthetisiert, die vom Statistischen Bundesamt (Destatis) und den Statistischen Ämtern der Bundesländer veröffentlicht wurden. Dieser Datensatz entstand in Zusammenarbeit mit Soofi, einem Konsortium deutscher KI-Forschungsinstitute und Start-ups, das souveräne Open-Source-Basismodelle für deutsche und europäische Unternehmen entwickelt.

Nemotron-Personas-Germany unterstützt deutsche Modellentwickler bei der Entwicklung souveräner KI-Systeme, die wichtige regionsspezifische, demografische, sprachliche und kulturelle Kontexte einbeziehen. Der Datensatz verbessert die Vielfalt bei der Generierung synthetischer Daten, mindert Verzerrungen in Daten und Modellen und trägt dazu bei, einen Modellkollaps (Leistungsabfall aufgrund unkontrollierten Trainings anhand der Ausgaben eines anderen Modells) zu verhindern, indem er die tatsächlichen geografischen und demografischen Verteilungen in Deutschland über alle 16 Bundesländer und 400 Kreise (Landkreise sowie kreisfreie Städte) hinweg widerspiegelt. Insbesondere ist der Datensatz so konzipiert, dass er die zugrunde liegenden demografischen Verteilungen entlang mehrerer Achsen repräsentativer abbildet als frühere Persona-Datensätze, darunter regionale Vielfalt (z. B. die Stadtstaaten Berlin, Hamburg und Bremen, die bevölkerungsreichen westlichen und südlichen Bundesländer sowie die östlichen Bundesländer), Einwanderungsgeschichte, Haushaltszusammensetzung, Bildungsniveaus und ein breites Spektrum an Berufen. So lassen sich beispielsweise hochwertige Chat-Konversationsdaten mit mehreren Gesprächsrunden erstellen, die realistische Namen, Altersangaben und Hintergründe enthalten und die authentische deutsche Lebensrealität widerspiegeln, wodurch diese Daten einzigartige Perspektiven und regionale Nuancen erhalten.

Der Datensatz wurde mit dem NeMo Data Designer erstellt, einem Compound-AI-System der Enterprise-Klasse zur Generierung synthetischer Daten. Er nutzt ein proprietäres probabilistisches grafisches Modell (PGM) in Verbindung mit einem unter der Apache-2.0-Lizenz stehenden google/gemma-4-31B-it-Modell sowie einer ständig wachsenden Anzahl von Validatoren und Evaluatoren, die in den Data Designer integriert sind. Eine erweiterte Version von Nemotron-Personas-Germany ist für die Verwendung im NeMo Data Designer verfügbar.

Der Datensatz gestattet die kommerzielle und nicht-kommerzielle Nutzung.

–

Nemotron-Personas-Germany is an open-source (CC BY 4.0) dataset of synthetically-generated personas grounded in real-world demographic and geographic trait distributions in Germany, to capture the diversity and richness of the population. It's the first German dataset of its kind aligned with statistics for names, sex, age, migration background, marital status, household type, education, employment, occupation, and location. All data is synthesized using official statistics from the 2022 German census (Zensus 2022) published by the Federal Statistical Office of Germany (Destatis) and the statistical offices of the Länder. This dataset was created in collaboration with Soofi, a consortium of German AI research institutes and startups that develops sovereign open-source base models for German and European companies.

Nemotron-Personas-Germany supports German model builders in developing Sovereign AI systems that incorporate important region-specific demographics, language, and cultural context. The dataset improves diversity of synthetic data generation, mitigates data and model biases, and helps prevent model collapse (degradation caused by uncurated training on another model's outputs) by reflecting Germany's real geographic and demographic distributions across all 16 federal states (Bundesländer) and 400 districts (Landkreise and kreisfreie Städte). In particular, the dataset is designed to be more representative of underlying demographic distributions along multiple axes, including regional diversity (e.g., the city-states of Berlin, Hamburg and Bremen, the populous western and southern states, and the eastern states), migration background, household composition, education levels, and a wide array of occupations, as compared to past persona datasets. As an example, one can produce high-quality multi-turn chat conversation data with realistic names, ages, and backgrounds that reflect the authentic German experience, bringing unique perspectives and regional nuances to that data.

Produced using NeMo Data Designer, an enterprise-grade compound AI system for synthetic data generation, the dataset leverages a proprietary Probabilistic Graphical Model (PGM) along with an Apache-2.0-licensed google/gemma-4-31B-it model and an ever-expanding set of validators and evaluators built into Data Designer. An extended version of Nemotron-Personas-Germany is available for use in NeMo Data Designer itself.

This dataset is ready for commercial or non-commercial uses.

Was NICHT im Datensatz enthalten ist (What is NOT in the dataset)

Da der Schwerpunkt auf Personas liegt, enthält der Datensatz keine weiteren Felder, die im NeMo Data Designer verfügbar sind, z. B. Vor- und Nachnamen, synthetische Adressen, Gesundheits- und Finanz-Personas sowie Big-Five-Persönlichkeitsmerkmale. Bitte kontaktieren Sie uns, um Anwendungsfälle für Unternehmen zu besprechen.

Alle Daten spiegeln zwar reale Verteilungen wider, sind aber vollständig künstlich erzeugt. Jede Ähnlichkeit von Namen oder Persona-Beschreibungen mit realen Personen, lebend oder verstorben, ist rein zufällig.

–

Given the emphasis on personas, the dataset excludes other fields available in NeMo Data Designer, e.g., first and last names, synthetic addresses, healthcare and finance personas, and Big Five personality traits. Please reach out to explore enterprise use-cases.

All data, while mirroring real-world distributions, is completely artificially generated. Any similarity in names or persona descriptions to actual persons, living or dead, is purely coincidental.

Datensatzersteller (Data Developer)

Soofi & NVIDIA Corporation

Veröffentlichungsdatum (Release Date)

Veröffentlicht auf Hugging Face am 20. Oktober 2026

–

Released on Hugging Face on October 20, 2026

https://huggingface.co/datasets/nvidia/Nemotron-Personas-Germany

Erstellungsdatum (Dataset Creation Date)

Oktober 2026

–

October 20, 2026

Lizenz und Nutzungsbedingungen (License/Terms of Use)

Dieser Datensatz unterliegt der Creative Commons Attribution 4.0 International License (CC BY 4.0).

–

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Anwendungsfälle (Use Case)

Entwicklerinnen und Entwickler, die an souveräner KI und am Training von LLMs arbeiten und/oder die Vielfalt synthetisch generierter Daten erhöhen, Verzerrungen in Daten und Modellen mindern und Modellkollaps verhindern wollen.

–

Developers working on Sovereign AI, training LLMs and/or looking to improve diversity of synthetically generated data, mitigate data/model biases, and prevent model collapse.

Datenversion (Data Version)

1.0 (20. Oktober 2026)

–

1.0 (October 20, 2026)

Vorgesehene Verwendung (Intended Use)

Der Datensatz Nemotron-Personas-Germany soll der Community dienen, offene Modelle weiter zu verbessern und den Stand der Technik voranzubringen. Die Daten dürfen frei zum Training beliebiger Modelle verwendet werden. Wir freuen uns über Rückmeldungen aus der Open-Source-Community und laden Entwicklerinnen und Entwickler, Forschende und Datenbegeisterte ein, den Datensatz zu erkunden und darauf aufzubauen.

Der Datensatz basiert auf den Verteilungen der selbst berichteten demografischen Daten des Zensus 2022. Sein Hauptziel ist es, die Entwicklung souveräner KI zu unterstützen, indem fehlende Daten und/oder mögliche Verzerrungen in heutigen Trainingsdaten ausgeglichen werden – insbesondere bei bestehenden Persona-Datensätzen, die für die Generierung synthetischer Daten verwendet werden. Trotz der verbesserten Datenvielfalt und der höheren Treue zur deutschen Bevölkerung sind wir weiterhin durch die Datenverfügbarkeit und eine vertretbare Modellkomplexität begrenzt. Daraus ergeben sich einige notwendige Unabhängigkeitsannahmen. Ebenso liegen umfassende Statistiken zur Geschlechtsidentität, unabhängig vom biologischen Geschlecht, in der amtlichen Volkszählung nicht in granularer Form vor. Weitere Verbesserungen überlassen wir künftigen Arbeiten. Namen sind eine statistische Näherung, kein Zensus.

Der Datensatz umfasst ausschließlich Erwachsene (ab 18 Jahren).

–

The Nemotron-Personas-Germany dataset is intended to be used by the community to continue to improve open models and push the state of the art. The data may be freely used to train any model. We welcome feedback from the open-source community and invite developers, researchers, and data enthusiasts to explore the dataset and build upon it.

The Nemotron-Personas-Germany dataset is grounded in distributions of self-reported demographic data in the 2022 German census (Zensus 2022). As such, its primary goal is to support Sovereign AI development by combating missing data and/or potential biases present in model training data today, especially when it comes to existing persona datasets used in synthetic data generation. Despite the improved data diversity and fidelity to Germany's population, we are still limited by data availability and reasonable model complexity. This results in some necessary independence assumptions. Similarly, comprehensive statistics on gender identity, independent of biological sex, are not fully available in a granular format from the national census. We leave further efforts to improve fidelity to future work. Names are a proxy, not a census.

Note that the dataset is focused on adults only (18+ years).

Details zum Datensatz (Dataset Details)

Der Datensatz enthält:

6 Mio. Personas in deutscher Sprache in 1 Mio. Datensätzen
27 Felder: 6 Persona-Felder, 6 Persona-Attributfelder, 14 Kontextfelder auf Basis amtlicher Bevölkerungs- und Arbeitsmarktstatistik und eine eindeutige Kennung
1,24 Mrd. Tokens insgesamt, davon 433 Mio. Persona-Tokens
16 Bundesländer und 400 Kreise (Landkreise und kreisfreie Städte)
775.000 eindeutige vollständige Namen (15.835 eindeutige Vornamen und 80.853 eindeutige Nachnamen)
12.000 Berufe (20.000 geschlechtsspezifische Berufsbezeichnungen) in 1.243 Berufsgattungen (KldB 2010), alle auf Basis realer Verteilungen
Umfassende Abdeckung der demografischen und geografischen Achsen
Verschiedene Persona-Typen: Beruf, Sport, Kunst, Reisen und Kulinarik sowie eine kurze Allzweck-Persona
Persona-Attribute in natürlicher Sprache: kultureller Hintergrund, Fähigkeiten und Fachwissen, Ziele und Ambitionen, Hobbys und Interessen

Nemotron-Personas-Germany wurde so konzipiert, dass es an der amtlichen Bevölkerungs- und Arbeitsmarktstatistik Deutschlands (Zensus 2022) ausgerichtet ist und diese um Bereiche erweitert, die für das KI-Training wichtig sind. Konkret bedeutet das:

Feingliedrige geografische Verteilung: Die Datensätze sind in allen 400 Kreisen (Landkreise und kreisfreie Städte) der 16 Bundesländer verankert und bilden die reale Bevölkerungsdichte und -verteilung in Deutschland ab.
Namensgebung: Auf Basis von Wikidata (P735/P734 × Geburtsjahr) und stefan-it/d-info-2005-names fließen 15.835 eindeutige Vornamen und 80.853 eindeutige Nachnamen in die Persona-Texte ein, die die sprachliche Vielfalt und die Herkunftsvielfalt der Bevölkerung in Deutschland widerspiegeln.
Standardisierter Berufsrahmen: Berufe und Branchen werden intern entlang der Klassifikation der Berufe 2010 (KldB 2010; 1.243 Berufsgattungen mit 20.000 Berufsbezeichnungen) und der Klassifikation der Wirtschaftszweige WZ 2008 gezogen, sodass die Erwerbstätigenstruktur im Datensatz der amtlichen Arbeitsmarktstatistik entspricht. Ausgeliefert werden die Berufsbezeichnung (occupation) und der Wirtschaftszweig (industry); die KldB-Codes selbst sind nicht Teil des Datensatzes.
Demografische Ausrichtung: Zentrale Merkmale wie Alter, Geschlecht, Familienstand, Haushaltstyp, Einwanderungsgeschichte und Bildungsniveau werden so gezogen, dass sie das demografische Profil Deutschlands im Jahr 2022 abbilden – eine statistisch repräsentative synthetische Bevölkerung.

Diese Erweiterungen bauen auf einer soliden Grundlage öffentlicher Daten auf und ermöglichen Personas, die statistisch fundiert und kulturell repräsentativ sind und dabei synthetisch, datenschutzfreundlich und offen bleiben.

–

The dataset contains:

6M personas in German across 1M records
27 fields: 6 persona fields, 6 persona-attribute fields, 14 contextual fields grounded in official demographic and labor statistics, and a unique identifier
1.24B tokens total, including 433M persona tokens
16 federal states (Bundesländer) and 400 districts (Landkreise and kreisfreie Städte)
775k unique full names (15,835 unique first names and 80,853 unique last names)
12k occupations (20k gender-specific job titles) across 1,243 occupation types (KldB 2010), all grounded in real-world distributions
Comprehensive coverage across demographic and geographic trait axes
A variety of persona types: professional, sports, arts, travel and culinary, plus a concise general-purpose persona
Natural language persona attributes: cultural background, skills & expertise, goals & ambitions, hobbies & interests

Nemotron-Personas-Germany was designed to align with Germany's official demographic and labor statistics (Zensus 2022), while extending them into areas important for AI training. In practice, this meant:

Granular Geographic Distribution: the dataset achieves high fidelity by anchoring records in all 400 districts (Landkreise and kreisfreie Städte) across the 16 federal states, mirroring the real population density and distribution across Germany.
Naming Conventions: by leveraging Wikidata (P735/P734 × birth year) and stefan-it/d-info-2005-names, the dataset incorporates 15,835 unique first names and 80,853 unique last names, reflecting the linguistic and ethnic diversity of the German population.
Standardized Occupational Framework: occupations and industries are sampled internally along the German Classification of Occupations 2010 (KldB 2010; 1,243 occupation types with 20k job titles) and the WZ 2008 classification of economic activities, so that the workforce distribution in the dataset matches official labor market statistics. The release ships the job title (occupation) and the economic sector (industry); the KldB codes themselves are not part of the dataset.
Demographic Alignment: key variables such as age, sex, marital status, household type, migration background and education levels are sampled to reflect the demographic profile of Germany from 2022, providing a statistically representative synthetic population.

Ausgangsdaten (Seed Data)

Um die soziodemografische und geografische Vielfalt und Komplexität der Bevölkerung Deutschlands abzubilden, nutzt Nemotron-Personas-Germany:

Daten der Volkszählung (Zensus 2022), veröffentlicht vom Statistischen Bundesamt (Destatis) und den Statistischen Ämtern der Länder
Die Klassifikation der Berufe 2010 (KldB 2010) der Bundesagentur für Arbeit, einschließlich geschlechtsspezifischer und geschlechtsneutraler Berufsbezeichnungen
Vor- und Nachnamen aus Namensvokabularen auf Basis von Wikidata (CC0; deutsche Vor- und Nachnamen nach Geschlecht und Geburtsjahrzehnt sowie Herkunftspools nach Staatsangehörigkeit), deren Häufigkeiten am D-Info-2005-Namenshäufigkeitsdatensatz kalibriert sind
Vornamen der ab 2000 geborenen Kohorte aus Geburtsregister-Daten nordrhein-westfälischer Kommunen über Open.NRW: Stadt Essen und Landeshauptstadt Düsseldorf (Datenlizenz Deutschland – Namensnennung – 2.0) sowie die Kommunen des Rhein-Erft-Kreises und des Kreises Düren über das Portal des KDVZ Rhein-Erft-Rur (CC BY 4.0)

–

In order to capture the socio-demographic and geographic diversity and complexity of Germany's population, Nemotron-Personas-Germany leveraged:

Population census data (Zensus 2022) published by the Federal Statistical Office of Germany (Destatis) and the statistical offices of the Länder
The German Classification of Occupations 2010 (KldB 2010) published by the Federal Employment Agency (Bundesagentur für Arbeit), including gender-specific and gender-neutral occupation titles
First and last names are drawn from name vocabularies built from Wikidata (CC0; German given names and surnames by sex and decade of birth, plus heritage pools by country of citizenship), with frequencies calibrated to the D-Info 2005 name-frequency dataset
First names of the cohort born from 2000 onward from municipal birth-register open data in North Rhine-Westphalia via Open.NRW: Stadt Essen and Landeshauptstadt Düsseldorf (Datenlizenz Deutschland – Namensnennung – 2.0) and the municipalities of the Rhein-Erft and Düren districts via the KDVZ Rhein-Erft-Rur portal (CC BY 4.0)

Schema

Der Datensatz umfasst 27 Felder: 6 Persona-Felder, 6 Persona-Attributfelder, 14 Kontextfelder und eine eindeutige Kennung, wie unten dargestellt. Der Datensatz enthält einen einzigen Split (train). Die zahlreichen Kontextfelder erlauben es Forschenden, gezielt bestimmte Bevölkerungsgruppen zu adressieren – was mit bestehenden Datensätzen nur schwer möglich ist. Die kategorialen Werte sind deutsch, wie im Zensus veröffentlicht.

Nemotron-Personas-Germany
  |-- uuid: string                        Global eindeutige Kennung
  |-- professional_persona: string        Berufliche Persona: Tätigkeitsfeld, zentrale Fähigkeiten und Eigenschaften
  |-- sports_persona: string              Sport-Persona: sportliche Interessen, Vereine, Einstellung zu Fitness
  |-- arts_persona: string                Kunst-Persona: Auseinandersetzung mit kreativem Ausdruck und Kunst
  |-- travel_persona: string              Reise-Persona: Reiseinteressen und -stil
  |-- culinary_persona: string            Kulinarische Persona: Essensvorlieben, Koch- und Essgewohnheiten
  |-- persona: string                     Kurze Allzweck-Persona, die das Wesen der Person erfasst
  |-- cultural_background: string         Beschreibung des kulturellen Hintergrunds der Person
  |-- skills_and_expertise: string        Berufliche und persönliche Fähigkeiten in Erzählform
  |-- skills_and_expertise_list: string   Liste der Fähigkeiten und Fachgebiete
  |-- hobbies_and_interests: string       Interessen und Freizeitaktivitäten in Erzählform
  |-- hobbies_and_interests_list: string  Liste der Hobbys und persönlichen Interessen
  |-- career_goals_and_ambitions: string  Berufliche Ziele und langfristige Karrierevorstellungen
  |-- sex: string                         Biologisches Geschlecht (Männlich, Weiblich)
  |-- age: integer                        Alter in Jahren (ab 18)
  |-- marital_status: string              Familienstand bzw. Status der eingetragenen Lebenspartnerschaft
  |-- migration_background: string        Mit / ohne Einwanderungsgeschichte
  |-- household_type: string              Typ des Haushalts, in dem die Person lebt
  |-- housing_tenure: string              Von Eigentümer/-in bewohnt oder vermietet
  |-- education_level: string             Höchster beruflicher oder akademischer Abschluss
  |-- employment_status: string           Erwerbstätig, erwerbslos, im Ruhestand, in Ausbildung oder sonstiges
  |-- job_position: string                Stellung im Beruf (Angestellte, Beamte, Selbstständige, mithelfende Familienangehörige) oder Nicht zutreffend
  |-- industry: string                    Wirtschaftszweig (WZ-2008-Abschnitt) oder Nicht zutreffend
  |-- occupation: string                  Aktuelle Berufsbezeichnung (geschlechtsspezifische Form) oder Nicht zutreffend
  |-- kreis: string                       Kreis (Landkreis oder kreisfreie Stadt)
  |-- bundesland: string                  Bundesland
  |-- country: string                     Land (Deutschland)

–

The dataset includes 27 fields: 6 persona fields, 6 persona-attribute fields, 14 contextual fields, and a unique identifier, as shown below. The dataset contains a single split (train). Researchers will find many contextual fields useful in zeroing in on specific demographics, which is challenging to do with existing datasets. Categorical values are in German, as published in the census.

Nemotron-Personas-Germany
  |-- uuid: string                        Globally unique identifier
  |-- professional_persona: string        Professional persona: field of work, key skills and traits
  |-- sports_persona: string              Sports persona: athletic interests, teams, approach to fitness
  |-- arts_persona: string                Arts persona: engagement with creative expression and the arts
  |-- travel_persona: string              Travel persona: travel interests and style
  |-- culinary_persona: string            Culinary persona: food preferences, cooking and dining habits
  |-- persona: string                     Concise general-purpose persona capturing the person's essence
  |-- cultural_background: string         Description of the person's cultural background
  |-- skills_and_expertise: string        Professional and personal skills in narrative format
  |-- skills_and_expertise_list: string   List of skills and areas of expertise
  |-- hobbies_and_interests: string       Interests and recreational activities in narrative format
  |-- hobbies_and_interests_list: string  List of hobbies and personal interests
  |-- career_goals_and_ambitions: string  Professional aspirations and long-term career objectives
  |-- sex: string                         Biological sex (Männlich, Weiblich)
  |-- age: integer                        Age in years (18+)
  |-- marital_status: string              Marital or registered-partnership status
  |-- migration_background: string        With / without migration history (Einwanderungsgeschichte)
  |-- household_type: string              Type of household the person lives in
  |-- housing_tenure: string              Owner-occupied or rented dwelling
  |-- education_level: string             Highest vocational or academic qualification
  |-- employment_status: string           Employed, unemployed, retired, in training, or other
  |-- job_position: string                Position in employment (employee, civil servant, self-employed, family worker) or Nicht zutreffend
  |-- industry: string                    Economic sector (WZ 2008 section) or Nicht zutreffend
  |-- occupation: string                  Current occupation title (gender-specific form) or Nicht zutreffend
  |-- kreis: string                       District (Landkreis or kreisfreie Stadt)
  |-- bundesland: string                  Federal state (Bundesland)
  |-- country: string                     Country (Deutschland)

Felder und Token-Anzahl (Field & Token Counts)

Der Datensatz umfasst 1,24 Mrd. Tokens (433 Mio. Persona-Tokens) in 1.000.000 Datensätzen und 26 Spalten, ohne die global eindeutige Kennung, gezählt mit dem Tokenizer von GPT-3.5-turbo (tiktoken cl100k_base). Die Daten decken alle 16 Bundesländer und 400 Kreise Deutschlands vollständig ab. Die folgende Tabelle zeigt für jede Spalte die Anzahl eindeutiger Werte, fehlender Werte und den Datentyp.

–

The dataset consists of 1.24B tokens (433M persona tokens) across 1,000,000 records and 26 columns, excluding the globally unique identifier, counted with the GPT-3.5-turbo tokenizer (tiktoken cl100k_base). Note that the data provides comprehensive coverage across all 16 federal states and 400 districts of Germany. The table below lists, for each column, the number of unique values, missing values and the data type.

column	num_unique	type	avg_tokens	total_tokens
`uuid`	1,000,000	string	N/A	N/A
`professional_persona`	1,000,000	string	80.688	80,688,492
`sports_persona`	1,000,000	string	72.208	72,208,266
`arts_persona`	1,000,000	string	71.316	71,316,334
`travel_persona`	1,000,000	string	72.704	72,703,800
`culinary_persona`	1,000,000	string	73.160	73,160,453
`persona`	1,000,000	string	63.188	63,188,437
`cultural_background`	1,000,000	string	146.047	146,046,995
`skills_and_expertise`	1,000,000	string	145.015	145,014,688
`skills_and_expertise_list`	999,952	string	75.365	75,364,887
`hobbies_and_interests`	1,000,000	string	137.551	137,551,252
`hobbies_and_interests_list`	999,999	string	74.790	74,790,278
`career_goals_and_ambitions`	1,000,000	string	133.342	133,341,541
`sex`	2	string	3.000	3,000,000
`age`	91	int64	1.000	1,000,000
`marital_status`	8	string	3.906	3,905,919
`migration_background`	2	string	8.798	8,798,357
`household_type`	5	string	8.024	8,024,190
`housing_tenure`	2	string	12.983	12,983,044
`education_level`	4	string	10.492	10,491,714
`employment_status`	5	string	4.682	4,681,826
`job_position`	6	string	7.618	7,617,787
`industry`	22	string	8.382	8,382,233
`occupation`	19,914	string	7.792	7,792,414
`kreis`	400	string	5.342	5,341,868
`bundesland`	16	string	4.703	4,703,219
`country`	1	string	3.000	3,000,000

Beschreibung des Datensatzes und Qualitätsbewertung (Dataset Description & Quality Assessment)

Die folgende Analyse schlüsselt den Datensatz entlang verschiedener Achsen auf, um die eingebaute Vielfalt und die Musterkomplexität der deutschen demografischen Daten zu verdeutlichen.

–

The analysis below provides a breakdown across various axes of the dataset to emphasize the built-in diversity and pattern complexity of the German demographic data.

Namen (Names)

Da der Schwerpunkt dieses Datensatzes auf Personas liegt, werden Namen nicht als eigene Felder bereitgestellt. In die Erstellung der Personas fließen jedoch 15.835 eindeutige Vornamen und 80.853 eindeutige Nachnamen ein, wodurch sich 775.000 eindeutige vollständige Namen ergeben. Das Namensvokabular stammt aus Wikidata (CC0); die Häufigkeiten sind an den Namenshäufigkeitsdatensatz D-Info 2005 (CC BY 4.0) angepasst. Die Namen werden unter Berücksichtigung von Geschlecht, Geburtskohorte, Staatsangehörigkeit, Kreis und Migrationshintergrund ausgewählt, sodass die Personas realistische Namensgebungsmuster in den verschiedenen Regionen und Gemeinden Deutschlands widerspiegeln.

–

Since the focus of this dataset is on personas, names aren't provided as dedicated fields. However, infused into persona generation are 15,835 unique first names and 80,853 unique last names, forming 775k unique full names. The name vocabulary comes from Wikidata (CC0); frequencies are calibrated to the D-Info 2005 name-frequency dataset (CC BY 4.0). Names are sampled conditional on sex, birth cohort, nationality, district and migration background, so that personas reflect realistic naming patterns across regions and communities in Germany.

Altersverteilung (Age Distribution)

Die Altersverteilung von Nemotron-Personas-Germany spiegelt die reale demografische Struktur des Landes wider. Die erwachsene Bevölkerung konzentriert sich auf die 50- bis 64-Jährigen – die geburtenstarken Jahrgänge der späten 1950er- und der 1960er-Jahre – mit einem Medianalter von 52 Jahren. Ab 65 nehmen die Zahlen stetig ab, mit einem sichtbaren Einschnitt bei den 75- bis 79-Jährigen, der die geburtenschwachen Jahrgänge am Ende des Zweiten Weltkriegs abbildet. Die Anteile von Männern und Frauen sind bis ins mittlere Alter nahezu gleich, während Frauen in den höheren Altersgruppen zunehmend in der Überzahl sind (1,6 Frauen je Mann ab 80), was die höhere Lebenserwartung von Frauen widerspiegelt.

Minderjährige unter 18 Jahren sind nicht im Datensatz enthalten.

–

The age distribution of Nemotron-Personas-Germany mirrors the country's real demographic structure. The adult population is concentrated between 50 and 64 years, reflecting the large baby-boomer cohorts born in the late 1950s and 1960s, with a median age of 52. After 65, counts decline steadily, with a visible dip at 75–79 that reflects the small birth cohorts at the end of the Second World War. Male and female shares are nearly equal through midlife, while women increasingly outnumber men in older age groups (1.6 women per man aged 80+), reflecting higher female longevity.

Note that minors under 18 are excluded from this dataset.

Familienstand nach Altersgruppe (Marital Status by Age Group)

Die Heatmap zeigt für jede Alterskohorte den Anteil der Personen nach Familienstand (ledig; verheiratet oder in eingetragener Lebenspartnerschaft; geschieden oder Lebenspartnerschaft aufgehoben; verwitwet oder Lebenspartner/-in verstorben; ohne Angabe – gemäß der Standardaggregation der Zensus-2022-Kategorien durch Destatis). Fast alle 18- bis 24-Jährigen sind ledig; ab Ende 30 wird „verheiratet“ zum häufigsten Status und erreicht bei den 60- bis 74-Jährigen mit rund zwei Dritteln den Höchstwert. Scheidungen sind im mittleren Alter am häufigsten (rund 16 % der 55- bis 64-Jährigen), während Verwitwung ab 70 stark zunimmt und ab Ende 80 dominiert (76 % bei 95+). All dies ist relevant, um authentische Lebenswelten und Personas im deutschen Kontext abzubilden.

–

The heatmap below displays the fraction of people for each age cohort by marital status (single; married or in a registered partnership; divorced or with a dissolved partnership; widowed or with a deceased partner; and not stated, following Destatis' standard aggregation of the Zensus 2022 categories). Almost all 18–24-year-olds are single, and marriage becomes the most common status from the late 30s, peaking at about two thirds of people aged 60–74. Divorce is most common in midlife (about 16% of those aged 55–64), while widowhood rises sharply after 70 and dominates from the late 80s onward (76% at 95+). All of these considerations are relevant to informing authentic life experiences and personas in the German context.

Haushaltstyp nach Altersgruppe (Household Type by Age Group)

Die Heatmap zeigt für jede Alterskohorte, in welchem Haushaltstyp die Personas leben (beschrieben wird der Haushalt, nicht die eigene Rolle der Persona darin). Die meisten 18- bis 19-Jährigen leben noch bei ihren Eltern: 48 % in einem Paarhaushalt mit Kindern und 37 % bei einem alleinerziehenden Elternteil. Paarhaushalte mit Kindern sind von Anfang 30 bis Anfang 50 erneut am häufigsten, wenn die Personas eigene Familien gründen. Ab Mitte 50 dominieren Paare ohne Kinder, und ab Mitte 80 werden Einpersonenhaushalte zum häufigsten Typ – rund zwei Drittel der über 90-Jährigen.

–

The heatmap below shows, for each age cohort, the type of household personas live in (note that this describes the household, not the persona's own role in it). Most 18–19-year-olds still live with their parents: 48% in a couple-with-children household and 37% with a single parent. Couple-with-children households are again the most common from the early 30s to the early 50s, as personas start their own families. From the mid-50s onward, couples without children dominate, and single-person households become the most common type from the mid-80s, reaching about two thirds of those aged 90+.

Bildungsniveau nach Altersgruppe (Education Level by Age Group)

Die Heatmap zeigt die Muster des Bildungsstands (höchster beruflicher oder akademischer Abschluss) über die Alterskohorten in Deutschland. Eine Lehre im dualen Ausbildungssystem ist in jeder Kohorte ab Mitte 20 der häufigste Abschluss, was die zentrale Rolle der beruflichen Bildung in Deutschland widerspiegelt; Fachschulabschlüsse (z. B. Meister oder Techniker) erreichen bei den 55- bis 64-Jährigen ihren Höchstwert. Hochschulabschlüsse sind bei den 25- bis 39-Jährigen am verbreitetsten (31–33 %) und nehmen in den älteren Kohorten stetig ab, in denen der Anteil ohne beruflichen Abschluss wieder steigt. Diese Muster zeigen den historischen Wandel des Bildungszugangs und veränderte gesellschaftliche Normen zu formaler Bildung.

–

The heatmap below captures patterns of educational attainment (highest vocational or academic qualification) across age cohorts in Germany. An apprenticeship in the dual vocational training system is the most common qualification in every cohort from the mid-20s onward, reflecting the central role of vocational training in Germany, while trade and technical school qualifications (e.g., master craftsman or technician) peak among 55–64-year-olds. University degrees are most prevalent among personas aged 25–39 (31–33%) and decline steadily in older cohorts, where the share without a vocational qualification rises again. These patterns highlight historical shifts in access to schooling and changing social norms regarding formal education.

Regionale Unterschiede im Bildungsstand (Geographic Intricacies of Education Attainment)

Dieser Ausschnitt zeigt, wie die Geografie den Bildungsstand und damit die Persona-Beschreibungen prägt. Die Choroplethenkarte zeigt für jeden Kreis den Anteil der Einwohner ab 25 Jahren mit Hochschulabschluss (Bachelor, Master, Diplom oder Promotion) und macht die deutlichen räumlichen Unterschiede zwischen Universitätsstädten und Metropolregionen einerseits und ländlichen Kreisen andererseits sichtbar – bezogen auf den Bundesdurchschnitt von 22,5 %. Universitätsstädte wie Heidelberg (47,0 %), Erlangen (45,9 %), München (45,3 %) und Freiburg im Breisgau (41,8 %) liegen vorn, während der ländliche Kreis Wittmund (7,7 %) sowie die Städte Delmenhorst (10,2 %) und Pirmasens (6,1 %) weit darunter liegen. Auf Länderebene führen die Stadtstaaten Berlin (35,5 %) und Hamburg (35,1 %); Sachsen-Anhalt (16,1 %) bildet das Schlusslicht.

–

This slice of the dataset demonstrates how geography informs education and therefore persona descriptions. The choropleth map shows, for each district (Kreis), the share of residents aged 25 and older who hold a university degree (Bachelor, Master, Diplom or doctorate), revealing marked spatial variation between university cities and metropolitan areas and rural districts, relative to the national average of 22.5%. University cities such as Heidelberg (47.0%), Erlangen (45.9%), Munich (45.3%) and Freiburg (41.8%) lead, while the rural district of Wittmund (7.7%) and the cities of Delmenhorst (10.2%) and Pirmasens (6.1%) sit far below it. At the state level, the city-states of Berlin (35.5%) and Hamburg (35.1%) rank highest and Saxony-Anhalt (16.1%) lowest.

Berufsgruppen (Occupational Categories)

Das Balkendiagramm zeigt die Verteilung der derzeit erwerbstätigen Personas (597.000, 59,7 % aller Datensätze) auf die großen Berufsgruppen. Diese beruhen auf den zehn Berufsbereichen der Klassifikation der Berufe 2010 (KldB 2010) und werden zusätzlich durch demografische und geografische Verteilungen geprägt. Die Abbildung zeigt die Gewichte der verschiedenen Beschäftigungsbereiche im Datensatz, etwa Gesundheit, Soziales, Lehre und Erziehung (21 %), Unternehmensorganisation, Buchhaltung, Recht und Verwaltung (18 %) sowie Rohstoffgewinnung, Produktion und Fertigung (18 %). Die rechte Tafel zeigt die Bildungsstruktur innerhalb der Bereiche: Hochschulabschlüsse dominieren in Sprach-, Literatur-, Geistes-, Gesellschafts- und Wirtschaftswissenschaften, Medien, Kunst und Kultur (59 %) sowie in Naturwissenschaft, Geografie und Informatik (51 %), während in Produktion, Bau, Verkehr und kaufmännischen Berufen rund die Hälfte der Beschäftigten eine Lehre hat.

–

The bar chart below reflects the professional distribution of the currently employed personas (597k, 59.7% of all records) across major occupation categories. These categories are grounded in the ten occupational areas (Berufsbereiche) of the German Classification of Occupations 2010 (KldB 2010) and are further informed by demographic and geographic trait distributions. This figure illustrates the prevalence of different workforce sectors in the dataset, such as health, social services, teaching and education (21%), business organisation, accounting, law and administration (18%), and production and manufacturing (18%). The right-hand panel shows the education mix within each area: university degrees dominate in the humanities, media and arts (59%) and in the natural sciences and IT (51%), while apprenticeships account for about half of all workers in production, construction, transport and commercial occupations.

Vielfalt der Personas (Persona Diversity)

Die genannten Attribute (und viele weitere) bestimmen letztlich die Vielfalt der generierten synthetischen Personas. Beispielhaft zeigt die folgende Analyse eine Vielzahl von Clustern in den beruflichen Persona-Beschreibungen. Sie wurden durch Clustering von Embeddings und Dimensionsreduktion auf 2D ermittelt: 20.000 zufällig ausgewählte berufliche Personas wurden mit einem mehrsprachigen Sentence-Transformer (paraphrase-multilingual-MiniLM-L12-v2) eingebettet, mit HDBSCAN geclustert und mit UMAP projiziert. Die acht größten Cluster sind mit ihrem dominierenden Berufsbereich und ihrer zentralsten Persona beschriftet (deutsches Original mit englischer maschineller Übersetzung).

–

The attributes above (and many more) ultimately affect the diversity of the synthetic personas being generated. As an example, the analysis below highlights a multitude of clusters within professional persona descriptions. These clusters are identified by clustering embeddings and reducing dimensionality to 2D: 20,000 randomly sampled professional personas were embedded with a multilingual sentence-transformer (paraphrase-multilingual-MiniLM-L12-v2), clustered with HDBSCAN and projected with UMAP. The eight largest clusters are labeled with their dominant occupational area and most central persona (German original with an English machine translation).

Verwendung (How to use it)

Der Datensatz lässt sich mit den folgenden Zeilen laden.

–

You can load the dataset with the following lines of code.

from datasets import load_dataset

# Deutsche Personas – German personas
nemotron_personas = load_dataset("nvidia/Nemotron-Personas-Germany", split="train")

Charakterisierung des Datensatzes (Dataset Characterization)

Methode der Datenerhebung (Data Collection Method)

Hybrid: menschlich, synthetisch, automatisiert (Hybrid: Manually-Collected, Synthetic, Automated)

Methode der Annotation (Labeling Method)

Nicht zutreffend (Not Applicable)

Datenformat (Dataset Format)

Zeichenketten und ganze Zahlen. Die vollständige Datenstruktur findet sich im Abschnitt „Schema“. (Strings and integers. The full data structure is provided in the Schema section.)

Umfang des Datensatzes (Dataset Quantification)

Anzahl der Datensätze: 1.000.000 Datensätze, 6.000.000 Personas (Record counts: 1,000,000 records, 6,000,000 personas)
Gesamtspeicher: 1,9 GB in 10 Parquet-Shards (Total data storage: 1.9 GB in 10 Parquet shards)

Ethische Überlegungen (Ethical Considerations)

NVIDIA ist überzeugt, dass vertrauenswürdige KI eine gemeinsame Verantwortung ist, und hat Richtlinien und Verfahren etabliert, die die Entwicklung eines breiten Spektrums von KI-Anwendungen ermöglichen. Beim Herunterladen oder bei der Nutzung gemäß unseren Nutzungsbedingungen sollten Entwicklerinnen und Entwickler gemeinsam mit ihren internen Teams sicherstellen, dass dieser Datensatz die Anforderungen der jeweiligen Branche und des Anwendungsfalls erfüllt und unvorhergesehener Missbrauch berücksichtigt wird.

Bitte melden Sie Qualitäts-, Risiko- und Sicherheitsprobleme oder sonstige Bedenken zu NVIDIA AI hier.

–

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Zitation (Citation)

Wenn Sie die Daten nützlich finden, zitieren Sie bitte:

–

If you find the data useful, please cite:

@software{nvidia/Nemotron-Personas-Germany,
  author = {Fromm, Michael and Khan, Abbas and Jude, Alex and Schweter, Stefan and Singh, Sarandeep and Habert, Severine and Sharabiani, Ashton and Jennings, Will and Meyer, Yev},
  title = {{Nemotron-Personas-Germany}: Synthetic Personas Aligned to Real-World Distributions for Germany},
  month = {October},
  year = {2026},
  url = {https://huggingface.co/datasets/nvidia/Nemotron-Personas-Germany}
}

systemd — Tue, 20 Oct 2026 21:16:56 GMT

SimReady Central capacity level

status.json sets the HF Space's capacity level: {"level": 0-4, "message"?: "..."}. 0 normal, 1 busy, 2 heavy load (vision checks paused), 3 very heavy (no query parsing, stills only), 4 read-only. The Space's UI reads it every minute.

systemd — Tue, 20 Oct 2026 21:16:53 GMT

Dataset Overview

Dataset Description:

The ICT and Artix PCBA USD Asset dataset contains board-specific OpenUSD packages for the ICT and Artix printed circuit board assemblies. The packages include assembled board geometry, component assets, and materials authored directly in OpenUSD. The USD assets are derived from the respective board CAD designs.

Both packages were generated from KiCad PCB sources and subsequently underwent manual mesh curation for AOI use. This included separating selected component-pin geometry from component bodies and removing board-logo geometry. Values unavailable from the source designs or component datasheets—including stencil, silkscreen, and plating thicknesses—are identified as derived values in each package’s provenance.json.

This dataset is ready for commercial or non-commercial uses.

Dataset Owner(s):

NVIDIA Corporation

NVIDIA Corporation owns the dataset packaging and derived OpenUSD assets. The source PCB designs are provided by Antmicro under the Apache License 2.0.

Dataset Creation Date:

September 28, 2026

Versioning:

1.0.0

Previous Version(s):

None

License/Terms of Use:

Use of the PCB USD dataset is governed by CC BY-4.0. The upstream PCB designs are licensed under Apache License 2.0. Each package includes the applicable upstream license and attribution notice.

Intended Usage:

Intended for demonstrating and evaluating the pcb-aoi synthetic-data workflow using two representative PCB assemblies. Users may render the ICT and Artix USD assets, introduce supported structural defects, and generate images and annotations for AOI development and evaluation. The assets are not intended to represent production inspection accuracy unless separately validated.

Dataset Characterization

Data Collection Method

[Synthetic] - The OpenUSD assets were generated from KiCad CAD sources and subsequently curated through manual mesh post-processing. Modifications included separating component pins from component-body meshes and removing board logos and their associated geometry. These changes improve scene structure and suitability for AOI synthetic-data workflows.

Labeling Method

[Hybrid: Manually-Labeled, Synthetic] - Component and semantic metadata is derived from the KiCad source data during conversion and supplemented by manually authored or reviewed prototype properties. Labels are stored as metadata on USD prims. No rendered images or separate image-label dataset are included.

Dataset Format

The dataset is distributed as two self-contained ZIP packages, one for each PCBA:

Artix DC-SCM
ICT Baseboard

Each package contains:

OpenUSD ASCII (.usda) files:
- One root board stage
- Board-geometry layers for substrate, copper, solder mask, pads, solder paste, silkscreen, holes, test points, and component placements
- Instanceable component-package prototype assets
JSON (.json) files containing package manifests and provenance metadata
Markdown (.md) documentation
Upstream license and attribution notices

Component identity, placement, physical properties, and provenance annotations are stored as metadata on USD prims.

The packages do not contain MDL files, external texture images, rendered images, or separate annotation files. Materials and display properties are authored directly in OpenUSD.

Dataset Quantification

The dataset contains two PCBA asset packages:

Artix DC-SCM
- Board dimensions: 90 × 120.405 × 1.6 mm
- 1 root OpenUSD stage
- 9 board-geometry USD layers
- 47 package-prototype USD files
- 499 placed component instances
- 12 identified test-point prims
- Uncompressed package size: approximately 512.5 MB
- ZIP distribution size: approximately 49.8 MB
ICT Baseboard
- Board dimensions: 150 × 50 × 1.59 mm
- 1 root OpenUSD stage
- 9 board-geometry USD layers
- 32 package-prototype USD files
- 203 placed component instances
- 12 labeled test points
- Uncompressed package size: approximately 321.0 MB
- ZIP distribution size: approximately 30.9 MB

Combined totals

2 PCBA board assets
2 root OpenUSD stages
18 board-geometry USD layers
79 package-prototype files representing 70 distinct prototype names; 9 prototypes are shared between the board packages
702 placed component instances
24 labeled test points
99 OpenUSD ASCII (.usda) files
111 uncompressed files, including USD assets, manifests, provenance records, notices, licenses, and README files
Approximately 833.5 MB uncompressed
Approximately 80.7 MB as two ZIP distributions

Component metadata provides up to 25 identity, placement, package, appearance, and provenance fields. These include reference designator, component type, package and footprint identifiers, manufacturer, manufacturer part number, nominal value, board side, nominal position and rotation, process class, pin count, material class, and provenance availability fields.

Additional information:

Component types:
- 305 capacitors
- 253 resistors
- 47 ICs
- 32 diodes
- 20 connectors
- 16 inductors
- 15 transistors
- 3 crystals
- 2 switches
- 9 “other”
Board side: 436 top-side and 266 bottom-side placements.
Assembly process: 679 SMT, 14 through-hole, and 9 mixed-process placements.

Reference(s):

Artix-dc-scm: https://github.com/antmicro/artix-dc-scm

Specific commit/version: https://github.com/antmicro/artix-dc-scm/commit/b9856f78179fce0739a4955dd4ed5f912619efe6

ICT-Baseboard: https://github.com/antmicro/ict-baseboard

Specific commit/version: https://github.com/antmicro/ict-baseboard/commit/181d6ed7d2d9dfb358dd2b64046e3debdf4ee9ac

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

systemd — Tue, 20 Oct 2026 21:16:34 GMT

Can VLMs judge animal cuteness? People vs. gemini-3.8-flash on 325 image pairs

We showed people and a VLM the same 325 pairs of pet images and asked both the same question: "Which animal is cuter?" Thirty people answered each pair through Rapidata, and google/gemini-3.8-flash answered each pair 30 times through OpenRouter. This dataset is the complete raw material of that comparison: the image pairs, all 9,750 human votes, and all 9,750 VLM votes with the VLM's confidence and stated reason for each one.

The write-up is on our blog: Can VLMs Judge Animal Cuteness?

The setup

Six image models made pictures for 22 prompts about cats, dogs and rabbits, like "a cute kitten curled up in a basket": gpt-image-2, seedream-5-lite, p-image, nano-banana-2, flux-2-flex and krea-2-medium. Pairing up each prompt's images gives 325 pairs. Neither people nor the VLM saw the prompt an image was made from.

People answered in the Rapidata app, which shows each person the two images in a random order:

The VLM got these system and user messages, with the two images in place of the placeholders:

[system]
You are shown two images, A and B. Answer the question by picking one of them. You must always pick A or B; ties are not an allowed answer.

[user]
Image A:
<first image>
Image B:
<second image>
Which animal is cuter?

Answer with the label of the image you prefer, a confidence between 0 and 1, and a reason of at most 15 words. Reply with JSON only.

Each image of a pair came first in 15 of the VLM's 30 calls, so a lean toward the first image cancels out.

How the two compare

	People	gemini-3.8-flash
Who answered	4,448 annotator sessions, 103 countries	one model
Votes per pair	30	30
Total votes	9,750	9,750
Pairs where all 30 votes agreed	0	216
Pairs split 16–14 or closer	53	8

People split 15–15 on the two kittens below, and the VLM picked the sleeping one all 30 times:

The VLM voted 30–0 on 216 pairs. On 163 of them most people picked the same image, and on 23 most people picked the other one. The other 30 were close to even for people, 16–14 or closer. Five of those pairs:

The VLM picked the image it was shown first in 52.7% of its calls. Flipping the order changed its majority pick on 31 of the 325 pairs.

Schema

One row per image pair.

column	description
`matchup_id`	The pair's id, e.g. `cat-01\|flux-2-flex\|krea-2-medium`
`prompt_id`	The generation prompt's id, e.g. `cat-01`
`generation_prompt`	The prompt both images were made from. Neither people nor the VLM saw it.
`model1`, `model2`	The image models that made `image1` and `image2`
`image1`, `image2`	The two images, embedded. The numbers are labels, not the order anyone saw them in.
`votes_image1_people`, `votes_image2_people`	People's votes per image (30 per pair)
`votes_image1_vlm`, `votes_image2_vlm`	The VLM's votes per image (30 per pair)
`detailed_results_people`	JSON list of every human vote: `votedFor` (`image1`/`image2`), `annotatorId`, `userScore`, `country`, `language`
`detailed_results_vlm`	JSON list of every VLM call: `votedFor`, `shownFirst` (`image1`/`image2`), `replicate` (0–29), `confidence`, `reason`, `inputTokens`, `outputTokens`, `costUsd`

import json

from datasets import load_dataset

ds = load_dataset("Rapidata/vlm-vs-crowd", split="train").remove_columns(["image1", "image2"])

# pairs where all 30 votes went to the same image
for side in ["people", "vlm"]:
    print(side, sum(min(r[f"votes_image1_{side}"], r[f"votes_image2_{side}"]) == 0 for r in ds))
# people 0, vlm 216

# how often the VLM picked the image it was shown first
calls = [c for r in ds for c in json.loads(r["detailed_results_vlm"])]
print(sum(c["votedFor"] == c["shownFirst"] for c in calls) / len(calls))  # 0.527

Notes on the data

Annotator ids are anonymized sequential ids (h0001, h0002, …). Votes with the same id came from the same annotator session, and the ids can't be linked back to Rapidata accounts. Each pair's 30 votes come from 30 different sessions. Three votes have no country. userScore is the annotator's Rapidata reliability score (the global userScore, from 0 to 1) on that vote.

The Rapidata app shows each person the two images in a random order, but which order a person saw isn't in this data. For the VLM, shownFirst records it. model1 and model2 come from the benchmark's pairing and aren't random: gpt-image-2 is model1 in 109 pairs and nano-banana-2 in none. So use shownFirst, not the image number, to study order effects.

The VLM run cost $21.54 through OpenRouter, about 0.2 cents per call.

systemd — Tue, 20 Oct 2026 21:16:02 GMT

Safety and Refusal Decisions — Jev 150K

150,000 validated typed decisions for modern AI systems. Prompt/response safety and category-presence decisions for refusal routing.

General-purpose dataset: this release contains no clinical or medical data. Jev outputs are model predictions, not ground truth.

Each row pairs an upstream reference label with a pinned jev-1.13.0 prediction, confidence, structured answer, and provenance. The flat Parquet layout works directly with Hugging Face datasets, DuckDB, Polars, or pandas.

What it is good for

safety classifiers and refusal routers
category-conditioned moderation evaluation
confidence calibration and failure mining

Release summary

Item	Value
Rows / unique row IDs	150,000
Unique decision groups	23,644
Unique upstream source rows	23,644
Train	145,824
Validation	2,014
Test	2,162
Jev model	`jev-1.13.0`
Jev input tokens	68,645,221
Jev output tokens	3,571,798
Recorded estimated input cost	$2.883099282
Rows with a Jev error	0

Quick start

from datasets import load_dataset

dataset = load_dataset("MaziyarPanahi/Jev-Safety-Refusal-Decisions-150K")
row = dataset["train"][0]

print(row["state_preview"])
print(row["question_text"])
print(row["jev_prediction"], row["jev_confidence"])

For a private repository, authenticate first with hf auth login. Once the repository is public, the same snippet works without authentication.

How to read a row

Need	Columns
Stable identity and split	`row_id`, `group_id`, `output_split`
Decision task	`priority`, `task_family`, `primitive`
Model input	`state_json`, `state_preview`, `question_text`, `criteria_json`
Upstream reference	`gold_label`, `gold_score`, `gold_json`, `label_source`
Jev result	`jev_prediction`, `jev_confidence`, `jev_probabilities_json`, `jev_answer_json`
Full receipt	`request_json`, `jev_raw_response_json`, token usage, latency, request ID
Provenance	source dataset, pinned revision, row ID, license, URL, transformation

Decision primitives are intentionally simple:

choice: one categorical label plus a probability distribution;
noul: a bounded value from 0 to 1;
score: a structured numeric assessment.

Task mix

Task family	Rows
`safety_category_presence`	117,554
`prompt_safety`	23,644
`response_safety`	8,802

Decision primitives

Primitive	Rows
`noul`	117,554
`choice`	32,446

Source mix

Source	Rows
`nvidia/Aegis-AI-Content-Safety-Dataset-2.0`	150,000

Every row retains the pinned source revision, source license, source row ID, transformation, group ID, original label, complete Jev request, complete raw response, normalized prediction, probability distribution or score, usage, and latency.

Descriptive agreement audit

Across the 150,000 rows with directly comparable source labels, Jev/source-label agreement is 82.17%. This is a behavior audit, not an accuracy or benchmark claim; upstream labels can be noisy or underspecified.

Task family	Evaluated	Agreement	High-level interpretation
`prompt_safety`	23,644	81.46%	Descriptive Jev/source-label agreement only
`response_safety`	8,802	82.15%	Descriptive Jev/source-label agreement only
`safety_category_presence`	117,554	82.31%	Descriptive Jev/source-label agreement only

High-confidence (jev_confidence >= 0.90) disagreements: 1,678.

Validation and splitting

The release uses repaired component-safe splits. Its independent audit verifies:

exactly 150,000 rows and unique row IDs;
one priority only: safety_refusal_router;
no decision group or upstream source row crosses splits;
nonempty model, prediction, answer, raw response, completion, and provenance fields;
zero jev_error values;
valid choice probability distributions and in-schema predictions;
matching normalized and raw Jev answers;
pinned source revisions and flat, viewer-friendly Parquet shards.

Important limitations

Contains potentially disturbing or unsafe text.
Not a complete safety policy and not suitable as a sole production guardrail.
Confidence values are uncalibrated model outputs.
Exact-request duplicates are retained where they reflect repeated or conflicting upstream evidence; consumers may aggregate or deduplicate by request_json, group_id, or source identity.
The card declares license: other: every row records source-specific license and provenance, and users are responsible for reviewing the applicable upstream and provider terms.
This is a general decision-modeling dataset, not a clinical dataset.

systemd — Tue, 20 Oct 2026 21:15:53 GMT

Scorpion toxin binder design: data release

Claude designed proteins de novo to bind five scorpion toxins: Cn2, Ts1, AaH II, AaH I, and LqhIII. Two contract research organizations measured binding by surface plasmon resonance (SPR): Adaptyv Bio and Twist Bioscience. Seventeen designs are binders: seven against Cn2, four against AaH II, four against AaH I, two against LqhIII, and none against Ts1. Report: De novo design of protein binders against scorpion toxins with Claude (https://www-cdn.anthropic.com/31ac578806cffd8b3d92d8affe6edc4c04d502e1.pdf). Blog post: https://www.anthropic.com/research/claude-scorpion-toxin-binder-design.

_{How the scorpion neurotoxin Cn2 acts on sodium channels, how the antibody LR neutralizes it, and how Cn2-B1 competed with LR for Cn2 (60 s, no sound). That Cn2-B1 should neutralize Cn2 as LR does is an expected outcome, not yet tested by patch-clamp recordings in cells.}

Start here

design.csv has one row per design: sequence, each lab's result and KD against each toxin, and binder_against. Designs go by eight-character identifiers (uuid), binders also by names (Cn2-B1). Load uuid as text: a spreadsheet turns 946282e8 into 9.46E+13.

Dataset viewer

The default subset is design (one row per design, 211 rows). Every other top-level table and every table in supporting_tables/ is a subset named after its file; the two indexes are sensorgrams_index and structures_index; each subset has one split, full. The per-design folders in sensorgrams/ and structures/ are files to browse or download, as is the film in assets/.

What is in the package

File or folder	Contents
`design.csv`	All 211 designs, with both labs' results and `binder_against`.
`in_silico_scores.csv`	Predicted scores, in the order of `design.csv`.
`Cn2_competition_Twist_Bioscience.csv`	Competition with the antibodies LR, RU1, and 9004G for Cn2: 16 designs and four controls.
`BVP_ELISA.csv`	Baculovirus particle (BVP) ELISA for polyreactivity: 51 designs.
`toxins.csv`	The five toxins: sequence, family, and design target.
`controls.csv`	The control antibodies at both labs: main screen and reverse BLI.
`sensorgrams/`	One folder per design and control: picture, traces, and kinetic fits of the main screen, and the follow-up assays where run, indexed in `INDEX.csv`.
`structures/`	One folder per design: design model and predicted structures, indexed in `INDEX.csv`.
`supporting_tables/`	Twelve tables: how each design was made and expressed; follow-up assays; per-seed scores, interfaces, and thresholds (list in `COLUMNS.md`).
`COLUMNS.md`	Every file and every column, explained.
`LICENSE.md`	CC BY 4.0.

Getting started

from datasets import load_dataset
import pandas as pd

designs = load_dataset("Anthropic/scorpion-toxin-binder-design", split="full")  # the default subset, design.csv
origin = load_dataset("Anthropic/scorpion-toxin-binder-design", "origin", split="full")  # any other subset by its name
table = pd.read_csv("hf://datasets/Anthropic/scorpion-toxin-binder-design/design.csv", dtype={"uuid": str})  # uuid as text

# the tables only
hf download Anthropic/scorpion-toxin-binder-design --repo-type dataset --include "*.csv" --exclude "sensorgrams/*/*"
# the two folders of one design (888c8f04 is Cn2-B1)
hf download Anthropic/scorpion-toxin-binder-design --repo-type dataset --include "*/888c8f04/*"

Terms, in the report's words

We define a hit as a result at one lab indicating that a design binds a toxin or an unrelated protein (CLEC12A or OVA; off-target binding). The design, or the design-toxin pair, is then a hit at that lab.
We define a binder of a toxin as a design with 1) a hit against the toxin at both labs, in every replicate, 2) every KD from the main screen at 1 µM or tighter, 3) a fitted maximum response (fitted Rmax) at Twist Bioscience within the theoretical Rmax (the largest response that one toxin molecule per binding site can give), and 4) no signal above background against CLEC12A in either of Twist Bioscience's two SPR runs. For LqhIII, measured only at Adaptyv Bio, a binder needed a hit there in every replicate, a KD of 1 µM or tighter, and criterion 4.

How to cite

Claude Science, Nathaniel Horwitz, James van Geelen, Kat Kajderowicz, Amir Shanehsazzadeh. De novo design of protein binders against scorpion toxins with Claude. October 20, 2026. https://www-cdn.anthropic.com/31ac578806cffd8b3d92d8affe6edc4c04d502e1.pdf.

License

CC BY 4.0 (see LICENSE.md).

systemd — Tue, 20 Oct 2026 21:15:51 GMT

GEOM-DRUGS for S23D

GEOM-DRUGS molecular conformers in the format read by S23D, the conformer generation model from the preprint A standard transformer and attention with linear biases for molecular conformer generation. It holds the training data and the test_1000 benchmark of the released models S23D-S (ibm-research/S23D.8m.v1) and S23D-B (ibm-research/S23D.25m.v1).

Download

The evaluation needs only the benchmark files (212 files, 0.62 GB), fetched here with checksums.sha256:

hf download ibm-research/S23D.GEOM.processed --repo-type dataset \
    --include "zarr/test_1000/*" --include "index/eval_index/test_1000/*" \
    --include checksums.sha256 --local-dir /abs/path/drugs_20
export SMILES23D_DATA=/abs/path/drugs_20

Then run s23dtest as described in the S23D README. The hf command comes with huggingface_hub, which smiles23d installs.

To train as well, download everything (589 data files, 54.7 GB):

hf download ibm-research/S23D.GEOM.processed --repo-type dataset --local-dir /abs/path/drugs_20

The same eval-only download from Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ibm-research/S23D.GEOM.processed",
    repo_type="dataset",
    allow_patterns=[
        "zarr/test_1000/*",
        "index/eval_index/test_1000/*",
        "checksums.sha256",
    ],
    local_dir="/abs/path/drugs_20",
)

Download to a fixed absolute path with --local-dir (local_dir), not into the Hugging Face cache: the cache path changes with every commit to this repo, and a restarted training run expects its data at the same path.

Checksums

checksums.sha256 lists the SHA-256 of every data file. After an eval-only download, check the benchmark files:

cd /abs/path/drugs_20
grep -E '  (zarr|index/eval_index)/test_1000/' checksums.sha256 | sha256sum --check --quiet

After a full download, check every file:

cd /abs/path/drugs_20
sha256sum --check --quiet checksums.sha256

Both checks print nothing and exit with status 0 when every file matches. On macOS use shasum -a 256 --check in place of sha256sum --check. The record decoder can unpickle object arrays, so only read a copy of this data that passes the check.

Layout

The repository root is the directory that SMILES23D_DATA points to:

drugs_20/
  train/                       data*.bin, index.json      4,103,418 conformers, one item each
  val/                         data*.bin, index.json      30,414 molecules, one item each
  test_1000/                   data*.bin, index.json      the 1,000 benchmark molecules
  zarr/test_1000/              .zarray, 0                 the 1,000 benchmark molecules, Zarr v2
  index/eval_index/test_1000/  index.json, chunk_index*   208,304 evaluation entries
  checksums.sha256

Folder	Files	Size	Read by
`train/`	343	46.8 GB	training
`val/`	32	5.8 GB	training, when validation is turned on
`test_1000/`	2	1.44 GB	nothing in the release configs
`zarr/test_1000/`	2	0.62 GB	evaluation
`index/eval_index/test_1000/`	210	0.12 MB	evaluation

train/, val/ and test_1000/ are litdata folders; each index.json lists the chunks and the record schema. val/ and test_1000/ are ordered by atom count.
test_1000/ holds the same 1,000 molecules (104,152 reference conformers) as zarr/test_1000/, in the training format.
Each entry of index/eval_index/test_1000 is the position of a molecule in zarr/test_1000, and a molecule with K reference conformers appears 2K times. The two folders only work as a pair.
The evaluation groups conformers by each record's smiles string, so keep the strings exactly as stored: do not re-key or canonicalise them.

Read the data with the smiles23d package. It was tested with smiles23d v1.0.0, litdata 0.2.30, zarr 3.0.10, posebusters 0.6.5 and RDKit 2026.3.6.

Evaluation

The S23D benchmark generates 2K conformers for a molecule with K reference conformers, with 300 diffusion steps, and scores heavy-atom RMSD after alignment with a coverage threshold of 0.75 Å and no chirality correction. A complete run on this data prints n rmsd calcs: 77763496 and writes 208,304 rows to output_0.jsonl. A lower count means molecules were skipped; because test_1000 is ordered by atom count, a short run skips the largest ones and overstates the metrics.

Results from the preprint for S23D-S-1/9 (S) and S23D-B-1/13 (S). Mean / median:

Model	Recall COV (%) ↑	Recall AMR (Å) ↓	Precision COV (%) ↑	Precision AMR (Å) ↓
S23D-S (`ibm-research/S23D.8m.v1`)	80.7 / 89.9	0.483 / 0.458	57.5 / 57.5	0.757 / 0.700
S23D-B (`ibm-research/S23D.25m.v1`)	84.5 / 91.7	0.420 / 0.387	62.5 / 63.4	0.690 / 0.626

Sampling is random, so a rerun will not repeat these numbers exactly.

Training

Training reads only train/, plus val/ when validation is turned on. The released models were trained on this data. Training is not seeded, so a new run gives a different model. The training configs read train/ in stored order; with litdata 0.2.30, 2 GPUs (+task.trainer.devices=2) and the batch_size and num_workers of the released configs, every run sees the records in the same order.

Provenance

GEOM (Axelrod and Gómez-Bombarelli, 2022): the GEOM-DRUGS conformer ensembles, from Harvard Dataverse, doi:10.7910/DVN/JNGTDF, under CC0 1.0.
GeoMol (Ganea et al., 2021): the train, validation and test split and the 1,000-molecule test set.
MCF (Wang et al., 2024): the processed GEOM-DRUGS archives, with at most 20 conformers per molecule in train and validation, all conformers in test, the smi keys, and precomputed eigenvectors (eig_vals, eig_vecs) that S23D does not read.
Conversion with the scripts in smiles23d/tools/preprocessing: conformers are grouped into molecules by the smi key, and conformers with a disconnected bond graph or at most one atom are dropped (in test_1000, 106,775 conformers become 104,152). A rebuild from the MCF archives reproduces this content, but not the record order.

License

This dataset is released under the Community Data License Agreement – Permissive, Version 2.0. The GEOM data it is built from is under CC0 1.0. If you use it, please cite GEOM, GeoMol, MCF and S23D.

Citation

@article{axelrod2022geom,
  title   = {{GEOM}, energy-annotated molecular conformations for property prediction and molecular generation},
  author  = {Axelrod, Simon and G{\'o}mez-Bombarelli, Rafael},
  journal = {Scientific Data},
  volume  = {9},
  number  = {1},
  pages   = {185},
  year    = {2022},
  doi     = {10.1038/s41597-022-01288-4}
}

@inproceedings{ganea2021geomol,
  title     = {{GeoMol}: Torsional geometric generation of molecular {3D} conformer ensembles},
  author    = {Ganea, Octavian-Eugen and Pattanaik, Lagnajit and Coley, Connor W. and Barzilay, Regina and Jensen, Klavs F. and Green, William H. and Jaakkola, Tommi S.},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {34},
  pages     = {13757--13769},
  year      = {2021}
}

@inproceedings{wang2024swallowing,
  title     = {Swallowing the Bitter Pill: Simplified Scalable Conformer Generation},
  author    = {Wang, Yuyang and Elhag, Ahmed A. A. and Jaitly, Navdeep and Susskind, Joshua M. and Bautista, Miguel {\'A}ngel},
  booktitle = {Proceedings of the 41st International Conference on Machine Learning},
  series    = {Proceedings of Machine Learning Research},
  volume    = {235},
  pages     = {50400--50418},
  year      = {2024}
}

@misc{gurev2025standard,
  title         = {A standard transformer and attention with linear biases for molecular conformer generation},
  author        = {Gurev, Viatcheslav and Rumbell, Timothy},
  year          = {2025},
  eprint        = {2506.19834},
  archivePrefix = {arXiv},
  primaryClass  = {q-bio.BM},
  url           = {https://arxiv.org/abs/2506.19834}
}

systemd — Tue, 20 Oct 2026 21:15:06 GMT

FairStress

Dataset for the paper "Asymmetrically Unfair Decision Making and Reasoning in LLMs".

FairStress tests LLM fairness as a decision process rather than a single answer. It includes 300 decision-making scenarios across 29 high-stakes domains (hiring, healthcare, law, housing, etc.), each posed with two candidates who differ by a protected-attribute contrast (race, gender, religion, disability, age, ...). Every scenario is instantiated at 4 degrees of demographic-identity signaling (implicit to explicit) and, at the robustness level, paired with one of 78 adversarial prefixes that argue for one candidate or the other.

FairStress gives the complete, full analysis of procedural bias — not just whether a model's final answer is correct, but whether it holds a correct decision, and a tied decision, equally robustly regardless of which group it favors.

About the work

Standard bias benchmarks (BBQ, StereoSet, CrowS-Pairs) score models on the answer alone. FairStress instead measures procedural fairness — whether the process that produced a decision is fair — at four levels: (1) the decision itself, (2) how firmly the model holds that decision under an adversarial prefix, (3) whether its stated reasoning is biased, and (4) whether its internal activations are. Each of the 300 base scenarios (two candidates, one domain, one decision) is rendered as a ground-truth version (one candidate is objectively stronger) and an ambiguous version (a true tie), across 33 minority/majority identity contrasts, 4 degrees of how explicitly that identity is signaled, and up to 78 adversarial prefixes arguing for one candidate or the other.

Headline statistics

Quantity	Value
Base scenarios	300
Domains	29
Demographic groups	31
Minority–majority contrasts	33, in 8 categories
Intervention degrees	4 (D0 no identity → D3 explicit label)
Adversarial prefixes	78, in 15 families, 3 classes

Domains (29 domains with base-scenario counts — every domain is instantiated across every applicable contrast/degree/prefix)

Domain	n	Domain	n
Hiring & Employment	20	Insurance Claims	10
Medical & Healthcare	18	Moral Tradeoffs	8
Legal / Criminal Justice	18	Environmental Justice	8
Social Services	14	Content Moderation	8
Financial Decisions	12	Education (Student-facing)	8
Disability Accommodation	10	Data Privacy	8
Mental Health	10	Consumer Protection	8
Immigration & Asylum	10	Legal Records	8
Education	10	Housing (Tenant-facing)	8
Public Safety	10	Epistemic Credibility	8
Veterans Benefits	10	Narrative / Personal Stmt.	6
Criminal Reentry	10
Housing & Tenancy	10
Academic Integrity	10
Child & Family	10
Child & Family (Application)	10
Workplace Discipline	10
Elder Care	10
Total			300

Contrast categories (33 contrasts, 31 demographic groups)

Category	n	Example contrasts
Race	4	Black/white, Hispanic/white, Arab/Hispanic, Arab/Black
Sex / gender	7	Black female/male, white trans woman/man, white trans man/man
Sexuality	2	gay/heterosexual, lesbian/heterosexual
Race × sex	7	Black female/white male, Hispanic trans woman/white male
Religion	8	Jewish/Muslim, Muslim/Christian, Sikh/Christian, Hindu/Muslim
Religion × gender	3	atheist male/Muslim female, Jewish male/Christian female
Age	1	older/younger worker
Disability	1	disabled/non-disabled
Total	33

Adversarial prefixes (78, 15 families, 3 classes)

Class	Families	n
Directional (matched pro-A/pro-B pairs)	authority, social consensus, representation, historical guilt, synthetic evidence, emotional appeal, legal liability, synthetic statistics, disadvantage claim	36
Neutral	evaluation awareness, stakes framing, reactance	18
Control	neutral-relevant, neutral-irrelevant, placebo	24
Total		78

Splits

Split	What it is
`full`	Recommended split to run. The complete generated corpus — every scenario × domain × signaling-degree × adversarial-prefix combination. This is the exact data behind the paper's results; running a model against it gives the complete, exact analysis of procedural bias.
`core`	FairStress-Core — a subset of `full`, selected via 2PL IRT (Fluid Benchmarking) rather than randomly. A lighter, cheaper option when running the full 13M-row corpus isn't feasible.

Usage

from datasets import load_dataset

# full corpus — recommended: the complete, exact analysis of procedural bias
full = load_dataset("PardisSzah/fairstress", split="full")

# FairStress-Core — lighter/cheaper subset
core = load_dataset("PardisSzah/fairstress", split="core")

systemd — Tue, 20 Oct 2026 21:14:53 GMT

OEIS Ladder 10B

A 10-billion-token pretraining corpus of integer sequences, built from the On-Line Encyclopedia of Integer Sequences (OEIS): every training-split OEIS sequence once, plus 9.2B tokens of distinct, transformed copies of OEIS sequences. It was used to pretrain NextTerm-2, a scaling ladder of byte-level language models (5M, 50M and 440M parameters) for a study of in-distribution vs out-of-distribution generalization in post-training.

Each row is one sequence written as comma-separated integers, e.g. {"id": "A000045_copy0", "text": "0,1,1,2,3,5,8,13,21,34,..."}.

Composition

	Rows	Tokens*	Share
Real OEIS sequences (`A……_copy0`)	370,869	754,629,901	7.5%
Generated, transformed sequences (`A……_gen…`)	12,544,402	9,245,370,098	92.5%
Train total	12,915,271	9,999,999,999
Validation (held-out OEIS sequences)	19,035	3.7M

*Tokens are byte-level training targets: the UTF-8 bytes of text plus one end-of-sequence token per row.

Rows are shuffled. The train split is in data/train-000NN-of-00010.jsonl (about 1 GB each); the held-out set is data/validation.jsonl.

How it was built

Real rows. The term list of each training-split OEIS sequence, taken from its b-file and truncated to a prefix of at most 4,096 bytes. Each appears exactly once.

Generated rows. Repeatedly:

Draw a source sequence from the real rows, with probability proportional to 1/√(its length in tokens). Uniform sampling spends most of the budget on very long sequences; 1/length over-samples short, idiosyncratic ones.
Transform it (aug_0):
- with probability 0.3 (sequences of ≥6 terms), one length-reducing transform: take every 2nd or 3rd term, first differences, or a prefix of ≥⅔ of the terms;
- then 0–2 length-preserving transforms, each chosen uniformly from: multiply by 2–5, negate, negate every other term, add a constant in [−10, 10], swap adjacent pairs, partial sums, reduce mod 2–10.
Cap at 4,096 bytes, and require at least 3 terms.
Reject exact duplicates of any real row or any earlier generated row, and draw again. 48.3% of draws were rejected; every generated row is distinct.

Generation statistics (sources drawn, rejection counts, seed) are in generation_manifest.json.

Validation split and overlap

The validation sequences are a random 5% of OEIS A-numbers, held out of the real rows and never used as augmentation sources, with exact duplicates of training sequences removed. They double as the OEIS-Eval-Neo benchmark (the 19,034 rows with at least two terms): predict the last term from the preceding ones.

OEIS is full of near-duplicates and transform relationships between entries (e.g. every-other-term of n⁵ is (2n)⁵), so held-out does not mean unseen. For 586 validation sequences (3.1%), some training row starts with the complete validation sequence: 102 via real OEIS entries that extend it, the rest via generated rows (only 11 are exact duplicates). Models score higher on those rows, but excluding them lowers OEIS-Eval-Neo by only 0.6–0.9 points for the NextTerm-2 models (1.0 for NextTerm-440M).

Models trained on it

The NextTerm-2 models: Qwen3-shaped byte-level models (vocabulary of 256 bytes + BOS/EOS/PAD), one pass over the train split, Muon + AdamW, cosine schedule, with identical hyperparameters at every size (Muon learning rate 2e-3, AdamW β2 0.99):

Model	Val loss (nats/byte)	OEIS-Eval-Neo	M1 macro MAPE	Ryskina & Knight
NextTerm-2-5M	1.134	23.4%	20.85	24/57
NextTerm-2-50M	0.960	31.6%	18.36	38/57
NextTerm-2-440M	0.881	36.3%	17.11	38/57

For reference, the M1 competition's Naive2 baseline scores 17.80 macro MAPE, and NextTerm-440M (trained on 14B tokens of a 1/length-sampled predecessor of this recipe) scores 34.4% on OEIS-Eval-Neo.

Loading

from datasets import load_dataset

train = load_dataset("N8Programs/oeis-ladder-10b", split="train", streaming=True)
val = load_dataset("N8Programs/oeis-ladder-10b", split="validation")

License and attribution

The sequences come from The On-Line Encyclopedia of Integer Sequences, published electronically at https://oeis.org, by The OEIS Foundation Inc., and are used under the Creative Commons Attribution-ShareAlike 4.0 license. The generated rows are adaptations of OEIS sequences and are shared under the same license.

systemd — Tue, 20 Oct 2026 21:14:52 GMT

MedSNIP: Building and Benchmarking Snippet-Level Granularity for Medical Fact Verification

Hasan Iqbal¹ · Sarfraz Ahmad¹ · Hyunjae Kim² · Sihyeon Park³ · Junjie Liao^4,5 · Qingyu Chen² · Preslav Nakov¹ · Yuxia Wang⁵

¹MBZUAI · ²Yale University · ³Korea University · ⁴Beijing Normal University · ⁵INSAIT, Sofia University "St. Kliment Ohridski"

A medical claim's correctness often depends not on the claim alone, but on the clinical structure around it. A claim may require a lab reference range, a causal or conditional link, or patient-specific details to be judged correctly, and atom-level decomposition can fragment these dependencies, leaving the verifier with clinically incomplete claims.

MedSNIP-Bench reformulates medical fact-checking around snippet-level verification, where clause-grouped units preserve local clinical structure. It covers 276 consumer-health and clinical-vignette responses, segmented by trained annotators into 2,524 snippets with dual in-general and in-patient-context labels and six structural pattern codes.

Loading

from datasets import load_dataset

ds = load_dataset("MBZUAI/MedSNIP")
ds["train"], ds["validation"], ds["test"]

Splits

Splits are made at the entry level, so all snippets from one response stay together.

Split	Entries	Snippets
train	180	1,599
validation	51	494
test	45	431
total	276	2,524

Fields

Field	Description
`snippet_id`	Stable identifier, `<entry_id>-S<n>`
`entry_id`	Source response this snippet came from
`query`	The original user question
`full_text`	The complete model response the snippet was cut from
`subset`	`consumer` (short patient questions) or `vignette` (clinical presentations)
`batch_id`	Annotation batch
`shared_context`	Entry-level context an annotator recorded, such as topic or patient details
`snippet_text`	The verification unit
`atoms`	The atomic claims grouped into this snippet
`pattern`	Structural pattern code, `A`–`F`
`is_ambiguous`	Annotator flag; all flagged snippets were adjudicated to binary labels
`notes`	Annotator rationale, where recorded
`label_atomic`	Label projected from the source physician atom labels
`label_human_general`	Annotator label, judged in general
`label_human_contextual`	Annotator label, judged for this patient
`split`	`train`, `dev`, or `test`

Because the corpus skews toward true snippets, the paper reports F1 on the false class as its primary metric.

Structural patterns

Each snippet carries one of six pattern codes. A, B and C merge multiple atoms; D, E and F keep an atom standalone.

Code	Pattern	Snippets
A	Enumeration over a shared subject	1,008
B	Causal or conditional chain	225
C	Conclusion plus its premises	321
D	Standalone claim	617
E	Distinct alternative, deliberately separated	39
F	Isolated claim kept apart from its neighbours	314

Of the 2,524 snippets, 1,554 group two or more atoms and 970 correspond to a single atom, over 5,838 atom references in total. 37 snippets carry the is_ambiguous flag.

Intended use

This dataset supports research on automated factuality evaluation for medical text generation. It is not a source of medical advice, diagnosis, or treatment recommendations, and should not be used for autonomous clinical decision support. Automated verification is imperfect: a verifier may accept false claims, reject correct ones, or abstain on clinically important cases.

Citation

@inproceedings{
anonymous2026medsnip,
title={Med{SNIP}: Building and Benchmarking Snippet-Level Granularity for Medical Fact Verification},
author={Anonymous},
booktitle={The 2026 Conference on Empirical Methods in Natural Language Processing},
year={2026},
url={https://openreview.net/forum?id=1vYUGLZnvr}
}

systemd — Tue, 20 Oct 2026 21:14:50 GMT

数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。

当前数据集卡片使用的是默认模版，数据集的贡献者未提供更加详细的数据集介绍，但是您可以通过如下GIT Clone命令，或者ModelScope SDK来下载数据集

下载方法

:modelscope-code[]{type="sdk"} :modelscope-code[]{type="git"}

systemd — Tue, 20 Oct 2026 21:14:45 GMT

T-DIG interactive test set

Prompts and diversity requests written by people using an image generator, collected for T-DIG: Text-conditioned Diversity for Image Generation (under review).

Each item is one prompt a participant typed, the four images the T2I model produced for it by independent sampling, and the participant's request for what should differ across the images (the axis), in their own words. These are the prompts and axes of the paper's test-set results and human evaluation.

T2I model	Items	Participants	First request was "generate"	Later requests
Qwen-Image-2512	105	21	75	29
Stable Diffusion 3.5 Large	101	20	62	52
Stable Diffusion 3.5 Medium	99	7	42	35
FLUX.2 [klein] 9B	101	24	71	53

The paper's automatic test-set metrics use the items whose first request was generate.

How it was collected

Participants recruited on Prolific used an interactive tool backed by one T2I model. They typed a prompt, saw four i.i.d. samples, and wrote in their own words what should vary across the set. They could then either generate a new set (generate) or edit the four images on screen (edit), and keep refining. Prompts and axes are verbatim, including spelling.

Fields

Field	Description
`item_id`	`<T2I model>-<number>`. Numbers have gaps where items were withheld (see Safety).
`renderer`	The T2I model that drew `iid_images`.
`prompt`	The prompt, as typed.
`axis`	The participant's first request for what should differ, as typed.
`mode`	`generate` (a new set from the prompt) or `edit` (edit the four images) for that first request.
`later_requests`	Further requests on the same prompt, in order: `axis`, `mode`, and `applied_to` (`original images` or `previous round`).
`participant`	Anonymous label (`p01`, `p02`, …), assigned per T2I model; labels are not linked across T2I models.
`source`	`interactive study (Prolific)`, or `interactive study (lab pilot)` for a few sessions from an earlier in-lab pilot.
`iid_images`	Paths of the four i.i.d. samples (1024×1024 PNG) the participant saw, relative to the T2I model's folder.

from datasets import load_dataset
ds = load_dataset("neulab/t-dig-testset", "qwen-image-2512", split="train")

Safety and privacy

Participants: no Prolific IDs, session IDs or other identifiers are included, only per-T2I-model labels.
Prompt text: scanned for e-mail addresses, links, phone numbers, handles and named public figures; none remain.
Images: every image was screened with a vision-language classifier for nudity, minors and violence, and the flags were reviewed by hand.
Withheld items: four items were removed and their numbers left as gaps:
- one depicting a named real person
- three depicting minors or young-looking people in underwear or swimwear
Remaining content: some items show mild fantasy or action violence (battle scenes, monsters) or adults in swimwear or revealing clothing. They were reviewed and kept.

Licenses

Text: prompts and axes are released under the license above.
Images: generated by Qwen-Image-2512, Stable Diffusion 3.5 Large and Medium, and FLUX.2 [klein]; their use is also subject to each model's license.

Citation

Citation information will be added after review.

systemd — Tue, 20 Oct 2026 21:14:33 GMT

StopLoss-Training-Data

Training data for the StopLoss monitors from StopLoss: Teaching Agents to Predict Failure and Stop Early. Each example is a window of a coding-agent run on a SWE-rebench task (GPT-5.6 Luna and Terra in Harbor's Terminus-2): the step count and the three most recent action–observation pairs, labelled with the run's final outcome.

🌐 Homepage | 📄 Paper | 💻 Code | 🤗 Models

Files

File	Contents
`train.jsonl`	training windows; keep the file order to reproduce the paper's training runs exactly
`validation.jsonl`	checkpoint-selection windows from held-out repositories

Fields

Field	Meaning
`task`	SWE-rebench task id
`k`	tool step at which the window ends
`prompt`	monitor input
`label`	`ABORT` if the run failed; for a successful run, `DELIVER` at its last two steps and `CONTINUE` before

Usage

From a clone of the GitHub repository:

hf download TIGER-Lab/StopLoss-Training-Data --repo-type dataset --local-dir data/stoploss
python training/train.py --recipe training/recipes/qwen2.5-7b-instruct.json \
    --data data/stoploss/train.jsonl --val-data data/stoploss/validation.jsonl \
    --out runs/qwen2.5-7b-instruct_s0 --seed 0

systemd — Tue, 20 Oct 2026 21:14:29 GMT

typed-extraction

A benchmark for typed extraction over a document: one piece of text in, a named map of typed fields asked at once, and for each field either a typed value or null when the document does not say.

It follows the System One extraction request shape (state, fields, reference_date; answers with value, span, p_null, confidence), so a case can be replayed against any API that implements it. It is the sibling of typed-decisions. It is an independent benchmark. It is not affiliated with any vendor of extraction models.

Leaderboard

test split: 1,690 cases, 17,733 fields. This table lists models scored zero-shot: they never saw train. Models fitted or trained on train are in the second table below; the two tables are not comparable.

Headline is exact match averaged over the 15 families; fields with a value is the same over fields whose gold is not null; computed is exact match on the 1,926 fields whose answer has to be worked out (totals, counts, day spans). Every model sees the same request: the document and the fields with their instructions, nothing else.

#	Model	Kind	Headline ↑	Fields with a value ↑	Hallucination ↓	Miss ↓	Computed ↑	`p_null` AUROC ↑
1	ScaleDown Extract	extraction API	0.782	0.723	0.117	0.200	0.570	–
2	Schematron V2 Small	extraction LLM	0.562	0.529	0.445	0.303	0.134	–
–	Always null (answers nothing)	reference	0.199	0.000	0.000	1.000	0.000	–
–	Regex rules	reference	0.129	0.152	0.993	0.000	0.026	0.503

Fitted or fine-tuned on `train`

These models were trained on this benchmark's train split, so they are not zero-shot and their scores are not comparable with the table above. The three test-only families never appear in train, so they still test transfer.

#	Model	Kind	Headline ↑	Fields with a value ↑	Hallucination ↓	Miss ↓	Computed ↑	`p_null` AUROC ↑
1	adaptive-classifier + MiniLM-L6	candidate classifier, trained on `train`	0.315	0.166	0.109	0.600	0.005	0.789
2	adaptive-classifier + bge-small	candidate classifier, trained on `train`	0.291	0.143	0.103	0.600	0.008	0.788
–	Always null (answers nothing)	reference	0.199	0.000	0.000	1.000	0.000	–
–	Regex rules	reference	0.129	0.152	0.993	0.000	0.026	0.503

Notes on the rows

ScaleDown Extract is the bar. It returns values that are not in the text (it does the arithmetic), and its weak spots are ExtractBench's long nested schemas and relative dates.
The adaptive-classifier baselines pick one of 40 candidate spans per field with a single yes/no classifier, so they rarely hallucinate but cannot compute, and for fields that have a value the right answer is among their candidates only 49% of the time. That ceiling, not the classifier, is what holds them back. They are published as reproducible open baselines, not as competitive extractors.
Always null answers nothing; it scores 0.199 because 31% of fields are absent. A model below it is doing more harm than good.
Remote models were called once, at temperature 0, with responses cached. LLMs got a JSON schema with one nullable property per field; documents longer than their context were cut (5 cases for Schematron), and that is counted.
Rows for LFM2.5 and Clairic will be added when their runs finish.

Scoring your model

pip install numpy pyarrow huggingface_hub
huggingface-cli download LocalLLaMA/typed-extraction --repo-type dataset --local-dir typed-extraction
python typed-extraction/score.py predictions.jsonl --by-family

predictions.jsonl has one line per test case, {"id": ..., "answers": {"<field>": {"value": ..., "p_null": ..., "confidence": ...}}}, with null (or [] for lists) for fields the document does not state. p_null and confidence are optional and enable the calibration metrics. Please report the headline, hallucination and miss together; any one alone can be gamed. To be added to the leaderboard, open a discussion on this dataset with the scorer's output and a link to the model, and say whether it was zero-shot or trained on train.

What is different about it

The gold is fact, not a teacher's opinion. Every answer is a human annotation from an open dataset, the record a constructed document was rendered from, or a Wikidata fact verified against the article text. No model labelled anything, so there is no teacher ceiling and frontier models can be ranked.
Absent fields are first-class. 31% of test fields have null gold, and the headline reports the hallucination rate (a value given for an absent field) separately from the miss rate.
Typed answers. Fields are text, number, date, time, datetime, duration or list. A value has to be the right type, not a string that happens to contain it. Numbers may be computed (totals, counts, a deposit stated as "two months' rent"), and dates may be relative ("delivered last Tuesday", resolved against reference_date).
Held-out families. Three constructed families (real-estate listings, meeting notes, medical appointment letters) appear only in the test split, so a score on them cannot come from memorised templates.
Variation. Constructed cases draw field names and instructions from synonyms. They shuffle field order, ask for a plausible field the document does not contain, and pad 30% of documents to 3–12k characters with decoy values of the same shape.

Splits

	test	train
cases	1,690	9,642
fields	17,733	74,014
null gold	31.1%	19.7%
held-out family fields	2,380	0

Test fields by family:

family	origin	cases	fields	what it tests
squad2	converted	200	1,908	spans; unanswerable questions as null
drop	converted	198	1,684	numbers, mostly computed
cuad	converted	102	4,182	41 clause lists per long contract (median ~26k chars)
extractbench	converted	30	2,115	nested schemas on long PDFs (10-Q, credit agreements, papers…), test only
wikipedia	distant	300	964	real articles, Wikidata values verified in the text
claim_email	constructed	80	617	relative dates, computed totals and day counts
invoice	constructed	80	681	due dates from payment terms, totals, line items
lease	constructed	80	770	end dates, deposits stated as multiples of rent
itinerary	constructed	80	608	datetimes, flight lists, journey minutes
shipping_notice	constructed	80	607	business-day ETAs, unit conversion
job_posting	constructed	80	609	salary ranges ("$85k–$110k"), skill lists
bank_statement	constructed	80	608	closing balances, largest debit, counts
real_estate_listing	constructed, held out	100	961	price per square foot, open-house datetimes
meeting_notes	constructed, held out	100	659	durations, "same time next week"
medical_appointment	constructed, held out	100	760	arrive-by times, absent copays and referrals

Test fields by type:

type	fields
list	6,854
text	5,278
number	3,480
date	1,119
duration	442
datetime	380
time	180

Present test gold is 8,192 verbatim spans, 2,020 normalised (written in another form), 1,926 computed and 86 relative.

Scoring

typed_extraction.metrics (numpy only) scores each field exact (0/1) and partially:

text: SQuAD-normalised exact match; partial credit is token F1.
number / duration: relative tolerance 1e-3, or rounding at the field's decimals.
date / time / datetime: equal ISO values.
list: multiset F1. A text item matches at token Jaccard ≥ 0.5 (CUAD's convention), and exact means F1 = 1.
null: right only when the prediction is null (an empty list for lists).

Headline: the mean over families, since families differ about 10× in field count. It is reported with hallucination rate, miss rate, p_null AUROC, ECE, and accuracy at 90% coverage (answering only the most confident 90%) for predictors that report confidence.

Gold quality

200 test fields were hand-checked against their documents, weighted towards null and derived gold. 2 were wrong (1.0%, 95% CI 0.3–3.6%). Both were fixed, together with every recurring issue the audit found:

CUAD exact clause boundaries, by moving to Jaccard list matching;
honorific-only differences;
ExtractBench instructions that asked for 'NA' where the gold is null;
ambiguous slash dates;
two malformed ExtractBench fields.

Null gold had no errors in the sample.

Known limitations:

The Wikipedia slice is thin: mostly 3 fields per case, and its few nulls are all death dates.
ExtractBench text comes from pdftotext. 5 of its 35 PDFs (image-only resumes) are excluded, and its LLM-judged metrics are not reproduced, so its scores are not comparable to the ExtractBench paper.
SQuAD 2.0 and DROP are in most models' pretraining data.
Itinerary durations use local clock times; the instructions say so.

Sources and licences

source	licence
SQuAD 2.0	CC BY-SA 4.0
DROP	CC BY-SA 4.0
CUAD v1	CC BY 4.0
ExtractBench, commit ca81763	MIT
Wikipedia text	CC BY-SA 4.0
Wikidata	CC0
constructed families	generated for this dataset

The dataset as a whole is released under CC BY-SA 4.0, as the share-alike sources require.

Format

Each row is one case:

column	content
`id`, `family`, `origin`, `held_out`	identifiers
`state`	the document
`reference_date`	the date relative expressions resolve against, if any
`fields`	JSON: name → `{type, instructions, options}`
`gold`	JSON: name → `{value, source, alts, spans}`

spans are [start, end) character offsets into state.

systemd — Tue, 20 Oct 2026 21:14:13 GMT

T-DIG policy training data

The GRPO training data of the T-DIG generation policies, from T-DIG: Text-conditioned Diversity for Image Generation (under review). These are the exact datasets the paper's released generation policies were trained and validated on, one per T2I model: Qwen-Image-2512, Stable Diffusion 3.5 Large, Stable Diffusion 3.5 Medium and FLUX.2 [klein] 9B. Code: github.com/neulab/t-dig. Policies: neulab/t-dig-checkpoints. Test set: neulab/t-dig-testset.

What a row is

A row gives the policy one base prompt and one axis, a request for what should differ across the images ("Monster creature type"), and asks it to write k = 4 prompts, one per image, that vary along the axis and otherwise keep the base prompt. Base prompts come from HPDv2; axes were proposed by a vision-language model that looked at four i.i.d. images of the prompt. An overall row has the axis overall and asks for variety across attributes the base prompt leaves open.

The four configurations hold the same base prompts, axes and order. They differ in the T2I model named in instruction_renderer, the word cap the instruction states, and the per-row thresholds, which are measured on that T2I model's own i.i.d. images.

Splits

Split	File	Rows	Contents
`train`	`train.parquet`	76,089	19,765 base prompts, 1,975 `overall` rows; four passes over the prompts in training order
`validation`	`val.parquet`	758	Every axis and `overall` row of 200 held-out base prompts, disjoint from training
`validation_monitor`	`val-monitor.parquet`	224	Per held-out prompt, its lowest-slot axis and its `overall` rows; a subset of `validation`
`probe_contrastive`	`probe-contrastive.parquet`	64	Held-out hard-contrastive axes; never trained on

A hard-contrastive axis names something the images must exclude ("instead of", "rather than", an unpaired "without", a leading "Remove"); those rows are removed from training. The order of train is the training order: each pass holds one axis per prompt, rotating the axis slot from pass to pass.

Fields

prompt: the chat message given to the policy, [{"role": "user", "content": <instruction>}]. instruction.txt shows the instruction of the first row.
reward_model.ground_truth: a JSON string read by the reward:
- base_prompt, axis: the base prompt and the axis, verbatim.
- tpips_factor: the text that conditions the TPIPS diversity score; equal to axis.
- k: images per set, 4.
- align_thresh: mean VQAScore of the T2I model's four i.i.d. images of the base prompt against the base prompt. The fidelity penalty applies below it.
- quality_thresh: mean HPSv3 of the same images. The quality penalty applies below it.
- length_cap: word cap of each written prompt, the base prompt's word count plus 20 (21 to 70). The length penalty applies above it.
extra_info: provenance and bookkeeping, including prompt_id (SHA-256 of the base prompt), row_id, index (training position), axis_kind (axis or overall), axis_slot and pass_number, cluster_id (HPDv2 prompt cluster), concise_factor (the elicitor's short name of the axis), source_renderer (the T2I model whose images the elicitor saw), instruction_version (tdig-v4-2), the thresholds again, align_thresh_images (images averaged) and length_cap.

data/<T2I model>/manifest.json lists row counts, settings, the build steps and the SHA-256 of every file; the trainer checks these hashes.

Prompt pool

prompts/curated-pool.json holds the 23,020 HPDv2 base prompts the paper started from: the prompts that passed both LLM filter passes, clustered by MiniLM embedding (sentence-transformers/all-MiniLM-L6-v2, cosine, average linkage) and cut at distance 0.35, keeping one prompt per cluster, drawn at random (seed 42) from the more central half of its members. Each entry in kept has the prompt's id (SHA-256 of the prompt), prompt, cluster_id and the filter pass's reason and variation_axes; cluster_summary lists every cluster's members. The training rows come from 20,000 of these prompts, 5,000 per T2I model. The code's iid.curated prepare --input prompts/curated-pool.json --allow-partial-input starts step (a) from this pool and reproduces the paper's i.i.d. images.

How it was built

The code repository's step (a) builds these files: HPDv2 prompt curation and clustering, four i.i.d. images per prompt per T2I model, axis elicitation, the axis export and a prompt-disjoint held-out split, then tdig.build_dataset --order rotated, tdig.iid_vqa attach, tdig.iid_quality attach and tdig.recap_dataset --relative-budget 20.

Training a policy with it

hf download neulab/t-dig-data --repo-type dataset --local-dir t-dig-data
T_DIG_SUBMIT="sbatch --nodes=2 --ntasks-per-node=1 --gpus-per-node=8 --time=72:00:00" \
$PY_TRAIN scripts/prepare_training_matrix.py --dataset t-dig-data/data/flux2-klein-9b \
  --out $T_DIG_RUNS/grpo-flux2-klein-9b --t2i-models flux2-klein-9b --sizes 9B \
  --tau 0.9 --lam 2.0 --learning-rate 2e-5 --metric vendi --length-weight 0.1 --quality-weight 0.05 --quality-margin 0 \
  --save-freq 50 --save-freq-early 25 --max-steps 2000 --validation offline \
  --track-metrics imagereward hpsv3 clip_score aesthetic --execute

These are the paper's settings. Then run scripts/preflight_training_matrix.py and scripts/submit_training_matrix.py as in the code repository's README.

License

CC BY 4.0. The base prompts are from HPDv2 (Apache-2.0) and remain under its license.

Citation

Citation information will be added after review.

systemd — Tue, 20 Oct 2026 21:13:43 GMT

systemd — Tue, 20 Oct 2026 21:13:42 GMT

CalebCool — Tue, 20 Oct 2026 21:01:24 GMT

数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。

当前数据集卡片使用的是默认模版，数据集的贡献者未提供更加详细的数据集介绍，但是您可以通过如下GIT Clone命令，或者ModelScope SDK来下载数据集

下载方法

:modelscope-code[]{type="sdk"} :modelscope-code[]{type="git"}

sibo0000 — Tue, 20 Oct 2026 17:19:47 GMT

数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。

当前数据集卡片使用的是默认模版，数据集的贡献者未提供更加详细的数据集介绍，但是您可以通过如下GIT Clone命令，或者ModelScope SDK来下载数据集

下载方法

:modelscope-code[]{type="sdk"} :modelscope-code[]{type="git"}

证据驱动的音视理解评测数据集

xiaoxing123321 — Tue, 20 Oct 2026 14:01:22 GMT

systemd — Sun, 18 Oct 2026 21:13:33 GMT

Xp-Teleop

Long-horizon humanoid manipulation on a Unitree G1 with Inspire FTP hands, recorded from a head-mounted RGB camera and cut into individual actions.

Xp-Teleop is the long-horizon loco-manipulation benchmark introduced in Ropedia W0.

27 episodes · 543 action segments · 96,064 frames · 30 fps · 1280×720

Layout

meta/
  info.json            schema for every field
  takes.jsonl          27 episodes
  segments.jsonl       543 action segments — the index
  tasks.jsonl          task list
  calibration.jsonl    camera intrinsics and extrinsics per episode
takes/<take_id>/segments/<segment_id>/
  video_rgb.mp4        clean RGB, without burned-in annotations
  video_keypoints.mp4  RGB with skeleton/keypoint overlays
  frame_timestamps.csv original per-frame timestamps
  frames.npz           per-frame arrays (see meta/info.json)
  masks.zip            robot body mask, one PNG per frame
  imu.csv              head IMU, ~1.6 kHz
  tactile.jsonl        Inspire tactile, 1062 taxels per hand
  caption.json         the action annotation

Load

import json, numpy as np

segments = [json.loads(l) for l in open("meta/segments.jsonl")]
s = segments[0]
d = np.load(f"{s['path']}/frames.npz", allow_pickle=True)

d["uv"]       # (n, 51, 2) keypoints in pixels
d["xyz_cam"]  # (n, 51, 3) keypoints in metres, camera frame
d["cam_c2w"]  # (n, 4, 4)  camera-to-world
d["q"]        # (n, 35)    joint angles

Clean RGB video

video_rgb.mp4 in every segment is encoded directly from the original JPEG frames (1280×720, 30 fps, H.264, CRF 18), without burned-in annotations; frame n corresponds to row n of frames.npz. frame_timestamps.csv gives the original per-frame timestamps (frame_idx, monotonic_ns, device_ts_usec) in the same order. video_keypoints.mp4 is the same footage with the robot skeleton rendered on top. The indices meta/raw_video_segments.jsonl and meta/raw_video_sources.jsonl list the source take and frame range of every segment.

Notes

frame_timestamps.csv also carries color_path / depth_path, the relative paths of the original colour JPEG and depth frames in the recording; those raw files are not part of this dataset.
In pour_and_wash_the_plate, segments c026–c032 have a video_keypoints.mp4 rendered from an earlier version of the skeleton; frames.npz and video_rgb.mp4 are the reference.

License

CC BY 4.0.

Citation

@article{liu2026ropedia,
  title={Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience},
  author={Liu, Tianqi and Zhao, Hanfei and Liang, Qifeng and Pang, Hui En and Xu, Jiarui and Mu, Sayang and Qiu, Penghao and Zhang, YingSheng and Wang, Jun and Cao, Ziang and Sun, Mingze and Cao, Yukang and Hong, Fangzhou and Chen, Zhaoxi and Liu, Ziwei},
  journal={arXiv preprint arXiv:2610.26855},
  year={2026}
}

kuangwang — Sun, 18 Oct 2026 11:45:01 GMT

Complete locally processed SeamlessInteraction full-duplex dataset: 17,318 conversations with turn, backchannel, overlap and pause annotations

数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。

当前数据集卡片使用的是默认模版，数据集的贡献者未提供更加详细的数据集介绍，但是您可以通过如下GIT Clone命令，或者ModelScope SDK来下载数据集

下载方法

:modelscope-code[]{type="sdk"} :modelscope-code[]{type="git"}

yyyyds139 — Sun, 18 Oct 2026 10:12:24 GMT

数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。

当前数据集卡片使用的是默认模版，数据集的贡献者未提供更加详细的数据集介绍，但是您可以通过如下GIT Clone命令，或者ModelScope SDK来下载数据集

下载方法

:modelscope-code[]{type="sdk"} :modelscope-code[]{type="git"}

DennisWeng — Fri, 16 Oct 2026 08:18:55 GMT

RoboFlywheel-Rigid — v1.3 Release

RoboFlywheel-Rigid is a simulation-ready rigid-object library for robot learning (grasping, manipulation, tabletop rearrangement, physical-reasoning benchmarks). This is the full v1.3 release: 524,564 assets across 30 domains and 1,207 populated classes (of a 1,212-class taxonomy) — cookware, kitchen prep, lab glassware, hand/power tools, fasteners, electronics, office, personal care, home decor, automotive, garden and toys.

Every asset here was processed end to end: identity & geometry QC → physics properties (mass, center of mass, full 3×3 inertia, friction) → convex collision decomposition → per-engine descriptions (URDF / MJCF / USD). All 524,564 assets in the tree are complete (_SUCCESS present, v13_status = ok, v13_usd_pass = 1).

Contents at a glance

Item	Value
Assets	524,564
Top-level domains	30 (`0` – `29`)
Populated classes	1,207 (out of a 1,212-class taxonomy)
Assets per class	3 – 15,335 (avg 434.6)
Physical parts per asset	1 – 71 (avg 5.1)
Collision hulls per asset	0 – 570 (avg 25.8)
Class provenance	357,117 kept original class · 147,573 reclassified into a better-matching class · 19,921 assigned to 7 newly defined classes
Engines	PyBullet · MuJoCo · Genesis · SAPIEN 3 · Isaac Sim

Every asset directory is self-contained with relative paths and is Z-up (content-level normalization — do not add an extra rotation at load time). The single source of truth for physics properties is part_physics.json (mass / com / 3×3 inertia / mu1 / mu2 / real_bbox / n_collision_hulls).

Light release: a class-balanced 10,000-asset subset of this package is published separately as the RoboFlywheel-Rigid light release (OSS release/light/, ModelScope dataset RoboFlywheel/Rigid). Leaf ids are shared — every light asset also exists here, with identical content.

Package layout

v1.3/
├── assets/<domain_id>/<class_id>/<leaf>/     # 524,564 asset directories
├── tables/
│   ├── master_table.parquet | .sqlite        # per-asset metadata (identity, geometry QC, physics, provenance)
│   ├── release_mapping.csv | .parquet        # original object_id → released leaf (524,611 rows)
│   └── taxonomy_v1.csv                       # 1,212-class taxonomy (domain/class ids + names)
└── category_mapping/<d>_<c>_<ClassName>/members.csv   # per-class member list, with provenance

Leaf directory names are <domain_id>_<class_id>_<index4>, e.g. assets/0/1/0_1_0000. release_mapping.csv is the join key back to the source pool: it maps the original object_id (pre-release identity) to new_leaf (this release), together with the redirect method used when an object was reclassified.

47 further entries in the tables did not make it into the tree (41 failed:no_source, 6 failed:no visual obj); they are reported in master_table with a null v13_usd_pass and have no directory. Treat the table as the complete record, the tree as the deliverable.

Per-asset layout

<leaf>/
├── m_<leaf>.urdf                # Primary description: visuals + all collision hulls inlined + 6-component inertia + textures (PyBullet/Genesis/SAPIEN)
├── coacd_multi.urdf             # Collision-mesh URDF (root/link_original, compatible with older downstreams)
├── mjcf/
│   ├── m_<leaf>.xml             # fullinertia + friction="mu1 mu2 1e-4" + textures (MuJoCo/Genesis first choice)
│   └── mesh/                    # self-contained: m_<leaf>.obj + material_0.png + m_<leaf>_collision_*.obj
├── mesh/                        # m_<leaf>.obj (Z-up) + <leaf>.obj (original) + m_<leaf>_collision.obj (merged hulls, same as v0.2) + material.mtl + material_0.png
├── part_physics.json
├── affordance/                  # semantic annotations (v1.0 original frame, not used in physics simulation)
├── <leaf>.usd                   # Isaac deliverable (baked: MassAPI / PhysicsMaterial / UsdUVTexture embedded)
└── _SUCCESS

Collision bodies: whole-object CoACD decomposition (threshold=0.05, max_convex_hull=64, ≤64 vertices per hull); parameters are recorded in part_physics.json.collision.

Common step: read the physics properties first

import json
pp = json.load(open(f"{asset_dir}/part_physics.json"))
mu1, mu2 = pp["whole_mu1"], pp["whole_mu2"]     # URDF has no friction field — pb/genesis/sapien must apply these at runtime

PyBullet — URDF + changeDynamics

import pybullet as p
p.connect(p.DIRECT)
bid = p.loadURDF(f"{asset_dir}/m_<leaf>.urdf")       # root link floats by default; pass useFixedBase=True to weld it
for l in range(-1, p.getNumJoints(bid)):
    p.changeDynamics(bid, l, lateralFriction=mu1, spinningFriction=mu2, rollingFriction=1e-4)
# Textures: the URDF <material><texture> is already inlined (effective when getVisualShapeData textureUniqueId >= 0)
# Note: collision AABB readings come out ~4-5mm larger than the hull mesh = Bullet's convex margin, not a data error

MuJoCo — MJCF preferred (full fidelity, native)

import mujoco
m = mujoco.MjModel.from_xml_path(f"{asset_dir}/mjcf/m_<leaf>.xml")   # friction / textures / fullinertia are native
# ⚠ Do NOT go through URDF: MuJoCo's URDF spec path ignores <visual> (no textures) and welds the root link into a static body

Genesis — URDF + explicit friction

import genesis as gs
gs.init(backend=gs.cpu)                              # only once per process
sc = gs.Scene(show_viewer=False)
ent = sc.add_entity(gs.morphs.URDF(file=f"{asset_dir}/m_<leaf>.urdf", fixed=True),
                    material=gs.materials.Rigid(friction=mu1))       # friction must be passed explicitly
sc.build()

SAPIEN 3 — URDF + PhysxMaterial per shape

import sapien, sapien.physx as px
scene = sapien.Scene()
loader = scene.create_urdf_loader()
# A single-link, jointless URDF is classified as an actor rather than an articulation:
# loader.load() always raises "URDF contains multiple objects";
# load_multiple returns (articulations, actors) — every asset in this library lands in actors
articulations, actors = loader.load_multiple(f"{asset_dir}/m_<leaf>.urdf")
a = actors[0]
mat = px.PhysxMaterial(mu1, mu1, 0.0)                 # static = dynamic = mu1; PhysX has no torsional friction field
rb = a.find_component_by_type(px.PhysxRigidBaseComponent)
for sh in rb.get_collision_shapes():
    sh.physical_material = mat                        # without this you silently get the 0.5 default and lose all friction

Isaac Sim — baked USD (the only reliable path)

from pxr import Usd
stage = Usd.Stage.Open(f"{asset_dir}/<leaf>.usd")     # or add_reference_to_stage
# RigidBody/MassAPI (mass+com+inertia principal axes)/per-hull CollisionAPI/PhysicsMaterial(mu1)/UsdUVTexture(sRGB) are all baked
# ⚠ Do NOT use the URDF/MJCF importer: external geometry stays in the importer's in-memory stage (active stage ends up with 0 meshes / 0 mass / 0 textures)

Notes and caveats

Friction does not come for free: the URDF standard has no friction field, and the PyBullet/Genesis/SAPIEN parsers all ignore it → you must read the values from part_physics.json and apply them at import time as shown above. MJCF and USD carry them natively, no extra work needed.
mu2 (torsional friction): only MuJoCo (2nd slot of friction) and PyBullet (spinningFriction) support it natively; Genesis/SAPIEN/Isaac only take mu1 (impact limited to spin damping).
Orientation: already normalized to Z-up at content level (visuals + hulls + com + inertia rotated together; recorded in part_physics.up_axis_fix).
real_bbox is a declared value, not a measured one (VLM/anchoring source; an individual axis can be off by 2-3×): use it only for Z-up orientation normalization. For gripper sizing or scene placement, read the measured AABB of the mesh or the hulls — not real_bbox.
affordance/: semantic annotations stay in the v1.0 original frame; if the asset was rotated (up_axis_fix.rotated non-empty), apply the same rotation to the annotations before use.
Isaac collision-body rendering: collision bodies are hidden in the USD with a cutout (opacity=0) while physics stays active; do not hide them via purpose/visibility (that turns physics off too).
USD materials: the visual material is OmniPBR (OmniPBR_Opacity.mdl sourceAsset, verified to sample textures correctly under Isaac RTX), and also carries a UsdPreviewSurface at preview purpose for standard USD tooling; non-Isaac viewers (e.g. usdview) go through the preview channel.
_SUCCESS semantics: marks a complete generation of this release (v1.3); it contains n_collision_hulls and the source v1.0 provenance line. A directory without _SUCCESS was not fully processed.

DennisWeng — Fri, 16 Oct 2026 08:17:14 GMT

数据集文件元信息以及数据文件，请浏览“数据集文件”页面获取。

当前数据集卡片使用的是默认模版，数据集的贡献者未提供更加详细的数据集介绍，但是您可以通过如下GIT Clone命令，或者ModelScope SDK来下载数据集

下载方法

:modelscope-code[]{type="sdk"} :modelscope-code[]{type="git"}

systemd — Thu, 15 Oct 2026 21:58:45 GMT

SenseNova RoboRSI Packages

Versioned experience and in-context learning (ICL) packages for benchmarks integrated into SenseNova RoboRSI. This dataset repository stores the resources consumed by the evaluation framework. Simulator environments and benchmark Docker images are distributed separately.

Available packages

Benchmark	Package version	Condition	Tasks	Package root
DexJoCo (`dexjoco`)	`v1.0.0`	`rand_full`	11	`dexjoco/v1.0.0/`
SIMPLE (`simple`)	`v1.0.0`	`simple_native_training_icl_v1`	5	`simple/v1.0.0/`
Bench2Dex (`bench2dex`)	`v1.0.0`	`task_icl_v3`	26	`bench2dex/v1.0.0/`
RoboDojo (`robodojo`)	`v1.0.0`	`public_training_frozen`	42 base / 54 mapped	`robodojo/v1.0.0/`
LIBERO-Long++ (`libero-long-plus-plus`)	`v1.0.0`	`efficiency_aligned_v12`	12 mapped / 12 packages	`libero-long-plus-plus/v1.0.0/`
LIBERO-Ultra-Long (`libero-ultra-long`)	`v1.0.0`	`efficiency_aligned_v12`	9 mapped / 9 packages	`libero-ultra-long/v1.0.0/`
LIBERO-PRO (`libero-pro`)	`v1.0.0`	`pro_persistent_memory_v18`	80 mapped / 40 packages	`libero-pro/v1.0.0/`
RoboCasa365 Target50 (`robocasa365`)	`v1.0.0`	`target50_same_task_seed0`	43 packages / 305 of 340 cells	`robocasa365/v1.0.0/`
BEHAVIOR (`behavior`)	`v1.0.0`	`official_21d_icl`	100	`behavior/v1.0.0/`

packages.json is the machine-readable release index. Paths in that index are relative to this repository root. Published version directories are immutable; changes to an experience package require a new version directory. Pin the full Hugging Face repository commit SHA when running evaluations.

Repository layout

README.md
LICENSE
packages.json
GUIDE.zh-CN.md
scripts/verify_package.py
dexjoco/v1.0.0/
  README.md
  LICENSE
  LICENSE-MIT
  THIRD_PARTY_NOTICES.md
  third_party/
  MANIFEST.sha256
  knowledge_packages.json
  rand_full/<task>/manifest.json
  rand_full/<task>/payload/...

The DexJoCo release preserves all 292 source files byte for byte (34,781,416 bytes). Its task directories contain 279 files (34,693,512 bytes), including 154 JPEG images and 27 NumPy files. All 181 image/NumPy files use Git LFS. The original MANIFEST.sha256 covers 291 package files, excluding itself. The catalog pins 11 task manifests and 268 payload files.

Download a pinned release

Install huggingface_hub. Authenticate with hf auth login if the repository is private and your account has access. Replace the placeholder below with a full commit SHA from this repository's history.

from pathlib import Path
from huggingface_hub import snapshot_download

download_root = Path(snapshot_download(
    repo_id="sensenova/SenseNova-RoboRSI-Packages",
    repo_type="dataset",
    revision="<FULL_COMMIT_SHA>",
    allow_patterns=["dexjoco/v1.0.0/*", "scripts/verify_package.py", "packages.json"],
    local_dir="./roborsi-packages",
))
package_root = download_root / "dexjoco" / "v1.0.0"
assert (package_root / "rand_full").is_dir()

huggingface_hub downloads the actual LFS file contents. For a Git checkout, install Git LFS and run git lfs pull to populate binary files.

Pass the package root, not the repository/download root, to the companion DexJoCo evaluator. From that evaluator's code checkout, for example:

python scripts/fetch_knowledge.py --local-dir /absolute/path/to/roborsi-packages/dexjoco/v1.0.0
python scripts/verify_release.py --require-knowledge

Alternatively, pass the same package root to run_frozen.py --knowledge-root. The archived package README describes the original standalone release layout; its direct fetch_knowledge.py --repo-id ... example must not be applied to this aggregate repository unless that downloader supports a package subdirectory. Use --local-dir as shown above. The catalog's knowledge/rand_full/... paths describe installation into the companion evaluator; the package-local path is rand_full/.... Keep frozen payload files and task manifests unchanged.

To verify the downloaded package, run the following from the same working directory as the download example. The trusted manifest hash below is the checksums_sha256 recorded for DexJoCo v1.0.0 in the release index:

python roborsi-packages/scripts/verify_package.py roborsi-packages/dexjoco/v1.0.0 --manifest-sha256 024dca42f2424cb77f9fdba433f6d6447fe9b8c6efac082febae0c7f3c9b7935

This checks hashes; it does not execute any bundled Python policy/helper code. Running the benchmark additionally requires the compatible evaluator and simulator environment. Download/hash verification is not an evaluation smoke test.

Provenance and licenses

DexJoCo simulator commit: 8d23b0fab23b17a58c4b55f3942e17013aaf8267. Source demonstration dataset: DexJoCo/DexJoCo-Datasets-Raw, revision 125df5c1019e97503929ef1a0ad8f90373436afa. See the package README, manifests, and per-task payload/SOURCES.md for preparation model, protocol, evidence, and source details. These packages apply to rand_full; do not relabel them as rand_obj packages.

Licenses are component-specific. Retain the package's LICENSE, LICENSE-MIT, THIRD_PARTY_NOTICES.md, third_party/, and per-task source records. In particular, the bundled notices identify GeoRT-derived material with CC BY-NC noncommercial terms. This aggregate repository does not relicense its contents.

Contribution guide

See the Chinese organization and upload guide, and the team's RoboRSI ICL material upload guide. Binary assets must be tracked with Git LFS. Benchmark IDs follow the benchmark Docker image naming convention, including dexjoco for DexJoCo. Each benchmark versions its packages independently; Docker image versions and package versions are separate identifiers and their compatibility must be documented.

SIMPLE v1.0.0

See the SIMPLE release guide for download, verification, task mapping, and integration details. Five frozen task packages are included: XMovePick, BendPick, Handover, Tabletop, and PickBetweenTables. XMoveBendPick has no ICL package in this version. These contain 2,385 frozen payload files plus their five original manifests (about 200 MB), including training references, images/videos, numerical trajectories, robot metadata, and helper scripts. No learned policy weights or evaluation episode logs are included.

Download simple/v1.0.0/*, packages.json, and scripts/verify_package.py at a fixed full repository commit. Use the SIMPLE entry's checksums_sha256 with the shared verifier. Package-local task roots are tasks/<Task>/; do not use DexJoCo-specific evaluator commands for SIMPLE. The archived package README is retained byte for byte; RELEASE.md contains the current publication guide. See SIMPLE component notices for provenance and data-license limitations. Hash verification does not run a simulator or model.

Bench2Dex v1.0.0

See the Bench2Dex release guide for pinned download, verification and RoboRSI Agent staging. The 26 frozen task ZIPs total 1.194 GB and contain 3,610 payload files. All 33 original release files, including hidden files and notices, are preserved byte for byte. The shared verifier supports this ZIP catalog without extracting or executing payloads; existing DexJoCo and SIMPLE directory catalogs remain supported.

Download bench2dex/v1.0.0/*, packages.json and scripts/verify_package.py at a full repository commit, then verify using the Bench2Dex index entry's checksums_sha256. Pass the version root to RoboRSI agent_bundle --source or run_suite --icl-source; this aggregate repository is not compatible with the standalone root-level --repo-id download example. Official benchmark assets and the simulator image are separate. The original pending-license notice remains applicable; this private submission does not establish public redistribution rights.

RoboDojo v1.0.0

See the RoboDojo usage guide. The release contains 27 complete sealed tar.gz experience packages: 21,387 original payload files and 2,521,855,428 uncompressed payload bytes, with 2,418,619,576 archive bytes. The catalog maps 42 original tasks and explicitly transfers base knowledge to 12 generalization variants; it includes no variant-specific demonstrations. These are historical public-training references, not weights or evaluation runs.

Download robodojo/v1.0.0/*, packages.json and scripts/verify_package.py at a full repository commit. Verify the version root using trusted manifest SHA256 7f65416beb515b888f253be2da8ae4a27fa868359d09200659000f4a97ef00d5. The shared verifier checks tar archive, package-manifest and payload hashes without extracting or executing the frozen helper code. The sensenova_roborsi XPolicyLab policy installs these exact archives and uses the standard Model/reset/update_obs/get_action and eval.sh entry points. Model service access, official simulator environment/Assets and task results are separate from package integrity. See the release's LICENSE_NOTICE.md for the absence of an independent experience-data license statement.

LIBERO Long++, Ultra-Long and PRO v1.0.0

The three LIBERO releases preserve the exact sealed packages used by the delivered no-VLA/no-SAM3 evaluation images: 12 Long++ v12 packages, 9 Ultra-Long v12 packages, and 40 PRO persistent-memory v18 packages mapped to 80 Task/Swap cases. Each release contains a frozen case map, machine-readable package catalog and complete checksum manifest. Download the selected <benchmark-id>/v1.0.0/* path plus scripts/verify_package.py at a full repository commit, then verify it with the checksums_sha256 value in packages.json. Simulator assets, model weights, credentials and evaluation results are not included.

RoboCasa365 Target50 v1.0.0

The release preserves the exact task memory built into the delivered HEART RoboCasa evaluation image ghcr.io/opensensenova/sensenova-roborsi-robocasa365@sha256:8494fef17ee5979a072967632a2cb6ca885a165e87b01fb88d40c9f88b5b82e1: 111 files from RLinf/RPent-memory robocasa/results/ at revision 551fc3157b3e56b40a3d3a3b4c7ff81721ebe89b, split into 43 per-task packages. Each cell uses only its own task's seed-0 memory; 7 tasks have none. Download robocasa365/v1.0.0/* and scripts/verify_package.py at a full repository commit and verify with the checksums_sha256 in packages.json. See the RoboCasa usage guide for mounting the merged memory root into the image. Simulator assets, VLA weights, credentials and evaluation results are not included.

BEHAVIOR v1.0.0

See the BEHAVIOR release guide for 100 task packages with 22,935 frozen payload files. All 100 source ZIP inventories were verified; retained files are byte-identical, while 693 generated bytecode files are omitted. Original metadata is retained outside the evaluation payload. Download behavior/v1.0.0/* at a full repository commit and verify it with the checksum in packages.json. The updated RoboRSI downloader uses sealed task directories directly; it does not accept the old task ZIPs. These are existing ICL materials, not new simulator/model results.

Scallions — Tue, 13 Oct 2026 06:47:19 GMT

Sea-ice observation / validation datasets (mirror)

数据进度与范围 / Status & Coverage

最后更新 (last updated): 2026-10-21 05:47 CST (UTC+8) 总进度 (overall): 164899/167353 (98.5%) 文件 (files), 已上传 (uploaded) 609.03 GB；82 个产品 (products)

分类说明见下文 Classification and folder layout。类型 (type): obs = 原始/L1–L2 观测；retrieval product = L3/L4 卫星反演产品；reanalysis/model = 再分析/模式；derived index/climatology = 派生指数/气候态。

Concentration / extent (密集度/范围)

产品文件夹 (product folder)	类型 (type)	级别 (level)	传感器/平台 (sensor)	方法 (method)	半球 (hemisphere)	时间 (time)	分辨率 (resolution)	大小 (size)	状态 (status)
`obs/satellite/osisaf_osi-450-a1/`	retrieval product	L4 (daily gridded, gap-filled CDR)	SMMR, SSM/I, SSMIS	Hybrid dynamic-tie-point PMW algorithm + RTM atmospheric correction	both	1978-10 – 2020-12	25 km EASE2, daily/monthly	31.28 GB	✅ 完成 (done) 28439/28439 (100.0%)
`obs/satellite/unibremen_amsr2-asi-6250/`	retrieval product	L3 (daily gridded)	AMSR2 (GCOM-W1)	ASI algorithm, 89 GHz polarisation difference	both	2012-07 – present	6.25 km polar stereographic, daily	18.43 GB	✅ 完成 (done) 10456/10456 (100.0%)
`derived-index/nsidc_g02135-v4/`	derived index/climatology	L4 index (extent/area tables + maps)	SMMR, SSM/I, SSMIS	NASA Team SIC → extent (≥15%) / area, daily & monthly, regional	both	1978-11 – present	25 km polar stereographic; tables	25.11 GB	✅ 完成 (done) 65866/65866 (100.0%)
`obs/satellite/esacci_sea-ice-concentration/`	retrieval product	L4 (daily gridded CDR)	SSM/I, SSMIS, AMSR-E, AMSR2	SICCI hybrid algorithm with dynamic tie-points + RTM correction	both	1991/2002 – 2020 (SSMI 12.5 km v3.0; AMSR 25/50 km v2.1 2002–2017)	12.5 / 25 / 50 km EASE2, daily	242.91 GB	✅ 完成 (done) 42361/42361 (100.0%)
`obs/satellite/nsidc_g02202-v6/`	retrieval product	L3 (daily/monthly gridded CDR)	SMMR, SSM/I, SSMIS	CDR = max of NASA Team and Bootstrap SIC	both	1978-10 – present	25 km polar stereographic, daily/monthly	8.85 GB	✅ 完成 (done) 2508/2508 (100.0%)
`obs/satellite/osisaf_osi-430-a/`	retrieval product	L4 (daily gridded ICDR)	SSMIS	OSI-450-a1 algorithm continued (interim CDR)	both	2021-01 – 2025-09	25 km EASE2, daily/monthly	3.87 GB	✅ 完成 (done) 3564/3564 (100.0%)
`obs/satellite/osisaf_osi-430-b/`	retrieval product	L4 (daily gridded ICDR, deprecated)	SSMIS	OSI-450 v2 algorithm continued (interim CDR)	both	2016 – 2023-05	25 km EASE2, daily	30.31 GB	✅ 完成 (done) 5348/5348 (100.0%)
`obs/satellite/concentration-extent/dtu_nimbus6-scams-v1.0/`	retrieval product	L3 (daily gridded)	Nimbus-6 SCAMS (22.2/31.7 GHz)	SIC and ice type from 22.235/31.65 GHz Tb	both	1975-06 – 1976-06	daily grid (see files)	2.09 GB / 2.09 GB	✅ 完成 (done) 2/2 (100.0%)
`reanalysis/concentration-extent/ncar_hadley-oi-sst-sic-v6.1/`	reanalysis/model	L4 analysis (model forcing)	HadISST1 (in situ + satellite) + NOAA OI.v2	Merged HadISST1/OI.v2 SST & SIC, anomaly-based blending (CAM AMIP forcing)	both (global)	1870-01 – 2022-06	1°, monthly	0.95 GB / 0.95 GB	✅ 完成 (done) 3/3 (100.0%)
`derived-index/concentration-extent/uhh_sie-ensemble-osi450a1/`	derived index/climatology	L4 index (extent ensemble)	SMMR, SSM/I, SSMIS (OSI-450-a1)	SIE (SIC≥15%) with 50-member correlated-noise ensemble	both (+regions)	1978 – 2020	daily regional totals (CSV)	0.17 GB / 0.17 GB	✅ 完成 (done) 25/25 (100.0%)
`derived-index/concentration-extent/uhh_sia-ensemble-osi450a/`	derived index/climatology	L4 index (area ensemble)	SMMR, SSM/I, SSMIS (OSI-450a)	SIA with 50-member correlated-noise ensemble	both (+regions)	1978 – 2020	daily regional totals (CSV)	0.20 GB / 0.20 GB	✅ 完成 (done) 25/25 (100.0%)
`obs/satellite/concentration-extent/dartmouth_is2-gridded-lif-fsd-v6/`	retrieval product	L3 (monthly gridded)	ICESat-2 ATLAS	Along-track ice/water class → linear ice fraction, floe size, waves	both	2018-10 – 2024-12	25 km, monthly	2.90 GB / 2.90 GB	✅ 完成 (done) 2/2 (100.0%)
`derived-index/concentration-extent/esr_antarctic-primary-ice-edge/`	derived index/climatology	L4 index (ice-edge contours)	SSM/I, AMSR-E, AMSR2	Northernmost 15% SIC contour (NT2/ARTIST), smoothed/unsmoothed	south	SSM/I – AMSR2 eras (see files)	daily contours (MATLAB)	1.99 GB / 1.99 GB	✅ 完成 (done) 5/5 (100.0%)
`obs/satellite/concentration-extent/snu_landsat8-sic-6.25km/`	retrieval product	L3 (gridded scenes)	Landsat-8 OLI	30 m ice/water classification aggregated to 6.25 km SIC	north (12 regions)	three years of Landsat-8 (see files)	6.25 km polar stereographic	0.27 GB / 0.27 GB	✅ 完成 (done) 12/12 (100.0%)
`obs/satellite/concentration-extent/caltech_modis-floe-segmentation-v2/`	retrieval product	L3 (daily floe maps)	MODIS Terra true colour	Floe segmentation → floe properties, SIC, floe-size distribution	north (Beaufort Sea)	2000 – 2022 (Mar–Sep)	256 m, daily	8.36 GB / 2.17 GB	✅ 完成 (done) 26/26 (100.0%)
`reanalysis/concentration-extent/uclouvain_antarctic-recon-1958-2023-v2/`	reanalysis/model	reconstruction (data assimilation)	9 ESM priors + station/ice-core data	Data-assimilation reconstruction, annual anomalies + uncertainty	south	1958 – 2023	annual fields	0.46 GB / 0.46 GB	✅ 完成 (done) 54/54 (100.0%)
`derived-index/concentration-extent/ucla_antarctic-sie-recon-20c/`	derived index/climatology	L4 index (reconstruction)	statistical reconstruction	Sector SIE anomaly reconstruction, 2500 members	south (sectors)	20th century	monthly (CSV)	0.41 GB / 0.41 GB	✅ 完成 (done) 6/6 (100.0%)
`derived-index/concentration-extent/uw_arctic-sie-recon-instrumental/`	derived index/climatology	L4 index (reconstruction)	CCSM4/MPI priors + HadCRUT/BE/GISTEMP	Data-assimilation reconstructions of total Arctic SIE	north	instrumental era (see file)	annual totals	0.01 GB / 0.01 GB	✅ 完成 (done) 1/1 (100.0%)
`obs/satellite/concentration-extent/dmi_msc-sic-v0/`	retrieval product	L4 (daily gap-free composite)	SMMR/SSMI/SSMIS, AMSR-E/AMSR2 + NIC/FMI/SMHI ice charts	Multi-source composite (OSI-450-a, SICCI-HR, OSI-458, charts), bias-corrected, gap-filled	both (global)	1982-01 – 2024-12	0.05° lat/lon, daily	0.00 GB / 141.30 GB	⏳ 进行中 (in progress) 0/5 (0.0%)

Thickness / freeboard / volume (厚度/干舷/体积)

产品文件夹 (product folder)	类型 (type)	级别 (level)	传感器/平台 (sensor)	方法 (method)	半球 (hemisphere)	时间 (time)	分辨率 (resolution)	大小 (size)	状态 (status)
`obs/satellite/awi_cs2smos/`	retrieval product	L4 (weekly merged)	CryoSat-2 SIRAL + SMOS MIRAS	Optimal interpolation of CS-2 radar-freeboard SIT and SMOS thin-ice SIT	north	2010-11 – 2025-04 (Oct–Apr)	25 km EASE2, weekly	4.61 GB	✅ 完成 (done) 2935/2935 (100.0%)
`reanalysis/psc_piomas-v2.1/`	reanalysis/model	model reanalysis	PIOMAS ice-ocean model (assimilates SIC, SST)	Coupled ice-ocean model with SIC/SST assimilation, NCEP forcing	north	1978 – present	PIOMAS grid 360×120, daily/monthly	2.43 GB	✅ 完成 (done) 449/449 (100.0%)
`obs/satellite/esacci_sea-ice-thickness/`	retrieval product	L3C (monthly gridded)	ERS-2 RA, Envisat RA-2, CryoSat-2 SIRAL	Radar freeboard + W99m/AMSR snow, hydrostatic equilibrium	both	ERS-2 1990s–2003, Envisat 2002–2012, CS-2 2010–2024	25 km EASE2 (NH) / 50 km (SH), monthly	0.84 GB	✅ 完成 (done) 504/504 (100.0%)
`obs/satellite/nsidc_atl20-v005/`	retrieval product	L3B (gridded)	ICESat-2 ATLAS	ATL10 laser total freeboard gridded daily/monthly	both	2018-10 – present	25 km polar stereographic, daily/monthly	0.58 GB	✅ 完成 (done) 186/186 (100.0%)
`obs/satellite/jpl_sits-gnssr-l3-antarctic-v1.0/`	retrieval product	L3 (7-day gridded)	Spire GNSS-R (grazing + near-nadir)	GNSS-R reflectivity → thickness, multi-observable combination	south	2022-10 – 2024-01	12.5 km EASE2 South, 7-day	1.27 GB	✅ 完成 (done) 134/134 (100.0%)
`obs/satellite/jpl_sits-gnssr-l3-arctic-v1.0/`	retrieval product	L3 (7-day gridded)	Spire GNSS-R (grazing + near-nadir)	GNSS-R reflectivity → thickness, multi-observable combination	north	2022-10 – 2024-01 (winters)	12.5 km EASE2 North, 7-day	3.67 GB	✅ 完成 (done) 307/307 (100.0%)
`obs/satellite/norce-tum_gnssr-freeboard-nh-v1.1/`	retrieval product	L3C (monthly gridded)	Spire grazing-angle GNSS-R	GNSS-R phase altimetry SSH → freeboard (experimental)	north	2023-10 – 2024-04	50 km EASE2 North, monthly	0.00 GB	✅ 完成 (done) 7/7 (100.0%)
`obs/satellite/thickness-freeboard-volume/legos_sit-cdr-ers-envisat-cs2-v2/`	retrieval product	L3 (monthly gridded)	ERS-1/2 RA, Envisat RA-2, CryoSat-2 SIRAL	Radar freeboard (CS-2 calibrated) + snow options, hydrostatic	both	1993/94 – 2023	monthly grid (see files)	6.51 GB / 6.20 GB	✅ 完成 (done) 74/74 (100.0%)
`derived-index/thickness-freeboard-volume/legos_sivol-cdr-v2/`	derived index/climatology	L4 index (integrated volume)	ERS-1/2, Envisat, CryoSat-2	Integrated sea ice volume from the LEGOS SIT CDR, regional, with uncertainty	both	1993 – 2023	monthly regional totals (CSV)	0.01 GB / 0.01 GB	✅ 完成 (done) 16/16 (100.0%)
`obs/satellite/thickness-freeboard-volume/legos_radar-freeboard-nh-v1.1/`	retrieval product	L3 (monthly gridded)	ERS-1/2 RA, Envisat RA-2, CryoSat-2 SIRAL	Radar freeboard; ERS/Envisat calibrated to CS-2 SAR TFMRA50	north	1993 – 2021	monthly grid (see files)	2.63 GB / 2.63 GB	✅ 完成 (done) 4/4 (100.0%)
`obs/satellite/thickness-freeboard-volume/legos_sit-cdr-sh-amsrsnow/`	retrieval product	L3 (monthly gridded)	ERS-1/2 RA, Envisat RA-2, CryoSat-2 SIRAL	Radar freeboard + SI-CCI AMSR snow climatology, hydrostatic	south	1994-04 – 2023-06	monthly grid (see files)	4.99 GB / 0.44 GB	✅ 完成 (done) 37/37 (100.0%)
`obs/satellite/thickness-freeboard-volume/gsfc_cs2wfa-antarctic-v0.1.1/`	retrieval product	L3 (monthly gridded)	CryoSat-2 SIRAL SAR/SARIn (Baseline-D)	Waveform fitting (CS2WFA) snow & ice freeboard, hydrostatic	south	2010-07 – 2021-08	25 km polar stereographic, monthly	0.28 GB / 0.28 GB	✅ 完成 (done) 134/134 (100.0%)
`obs/satellite/thickness-freeboard-volume/whu_arctic-sit-5km-v1.0/`	retrieval product	L3 (monthly gridded)	ERS-2 RA, Envisat RA-2, CryoSat-2 SIRAL	Lead detection, freeboard→SIT, inter-mission bias correction (DTU21 MSS)	north	1995-10 – 2025-04	5 km polar stereographic, monthly	1.70 GB / 1.69 GB	✅ 完成 (done) 1/1 (100.0%)
`obs/satellite/thickness-freeboard-volume/umd_is2smgpsit-v1/`	retrieval product	L4 (daily gridded, fused)	ICESat-2 ATLAS + SMOS/SMAP	Gaussian-process fusion of ICESat-2 SIT with SMOS/SMAP thin-ice SIT	north	2018-11 – 2025-04 (growth seasons)	25 km polar stereographic, daily	8.12 GB / 8.12 GB	✅ 完成 (done) 54/54 (100.0%)
`obs/satellite/thickness-freeboard-volume/umd_is2sit-summer-v1/`	retrieval product	L3 (monthly gridded)	ICESat-2 ATLAS (ATL10 R006)	ATL10 freeboard + SnowModel-LG snow (ERA5/MERRA-2), hydrostatic	north	2019–2021, May–Aug	25 km polar stereographic, monthly	0.27 GB / 0.27 GB	✅ 完成 (done) 33/33 (100.0%)
`obs/satellite/thickness-freeboard-volume/kochi-utokyo_nh-daily-sit-thermo/`	retrieval product	L4 (daily gridded)	AMSR-E, AMSR2	SIT from thermodynamic growth history of tracked PMW ice parcels	north	2007-01 – 2011-09, 2016-07 – 2025-12	daily grid (see files)	5.34 GB / 5.34 GB	✅ 完成 (done) 31/31 (100.0%)
`obs/satellite/thickness-freeboard-volume/uhh_smos-smap-sit-v1.0/`	retrieval product	L3 (daily gridded)	SMOS MIRAS + SMAP radiometer (L-band)	Combined 40° L-band Tb → thin-ice SIT (Tian-Kunze 2014)	north	2015-04 – 2018-04 (Oct–Apr)	12.5 km EASE2, daily	6.20 GB / 6.20 GB	✅ 完成 (done) 577/577 (100.0%)
`obs/satellite/thickness-freeboard-volume/lerma_pmw-sit-ssmi-v1.0/`	retrieval product	L3 (monthly gridded)	SSM/I, SSMIS	MLP trained on ICESat-2 SIT using 19/37 GHz Tb	north	1992 – 2020 (Oct–Mar)	grid per file	0.36 GB / 0.36 GB	✅ 完成 (done) 1/1 (100.0%)
`obs/satellite/thickness-freeboard-volume/norce_dasit-v100/`	retrieval product	L4 (drift-aware)	Envisat RA-2, CryoSat-2 + OSI SAF drift/SIC	CCI SIT parcels advected with PMW drift (drift-aware SIT)	north	Envisat 2002–2012, CS-2 2010–2020 (Oct–Apr)	grid per files	22.96 GB / 16.02 GB	✅ 完成 (done) 2/2 (100.0%)
`obs/satellite/thickness-freeboard-volume/ucl_daily-pan-arctic-sit/`	retrieval product	L4 (daily gridded)	CryoSat-2, Sentinel-3, ICESat-2	Daily-resolution SIT (Stroeve et al. 2024 method), several algorithms	north	2011 – 2022	daily grid (see files)	8.12 GB / 8.12 GB	✅ 完成 (done) 33/33 (100.0%)
`obs/satellite/thickness-freeboard-volume/tju_ice-3d-ml-sit/`	retrieval product	L4 (ML estimate)	not stated on record	Machine-learning SIT (ICE-3D); record has no description	north	2000 – 2023	not stated	4.88 GB / 4.88 GB	✅ 完成 (done) 24/24 (100.0%)
`obs/satellite/thickness-freeboard-volume/dmi_cs2-tfmra50-sit-2011-2022-v2/`	retrieval product	L3 (monthly gridded)	CryoSat-2 SIRAL (+ ICESat-2)	TFMRA50 radar-freeboard correction; snow depth & SIT retrieval	north	2011 – 2022 (winter)	monthly grid (see files)	0.04 GB / 0.04 GB	✅ 完成 (done) 1/1 (100.0%)
`obs/satellite/thickness-freeboard-volume/ucl_s3-narrs-v1.0.0/`	obs	L2 (along-track)	Sentinel-3A/B SRAL + MWR	NaRRS: MWR + SRAL waveform snow depth; SIT hydrostatic	north (50–81.35°N)	2016/17 – 2022/23 (Oct–Apr)	along-track, 23.5 km	4.06 GB / 1.43 GB	✅ 完成 (done) 2/2 (100.0%)
`obs/satellite/thickness-freeboard-volume/jpl_sits-gnssr-l2-antarctic-v1.0/`	obs	L2 (along-track)	Spire GNSS-R	GNSS-R reflectivity → along-track SIT	south	2022-10 – 2024-01	along-track	0.55 GB / 0.55 GB	✅ 完成 (done) 8/8 (100.0%)
`obs/satellite/thickness-freeboard-volume/jpl_sits-gnssr-l2-arctic-v1.0/`	obs	L2 (along-track)	Spire GNSS-R	GNSS-R reflectivity → along-track SIT	north	2022-10 – 2024-01	along-track	0.68 GB / 0.68 GB	✅ 完成 (done) 8/8 (100.0%)
`obs/satellite/thickness-freeboard-volume/norce-awi_dasit-75km-octapr/`	retrieval product	L4 (drift-aware parcels)	Envisat RA-2, CryoSat-2 + OSI SAF drift/SIC	DA-SIT customised: Oct–Apr advection, 75 km parcels	north	Envisat 2002–2012, CS-2 2010–2020	75 km parcels (GeoJSON)	14.51 GB / 14.51 GB	✅ 完成 (done) 20/20 (100.0%)
`reanalysis/thickness-freeboard-volume/nersc_topaz4-ml-sit/`	reanalysis/model	reanalysis + ML correction	TOPAZ4 (assimilates SIC, CS2SMOS) + ML	ML-corrected TOPAZ4 reanalysis thickness	north	1992 – 2022	TOPAZ4 grid (~12.5 km), daily	0.00 GB / 22.22 GB	⏳ 进行中 (in progress) 0/32 (0.0%)

Snow on sea ice (冰上积雪)

产品文件夹 (product folder)	类型 (type)	级别 (level)	传感器/平台 (sensor)	方法 (method)	半球 (hemisphere)	时间 (time)	分辨率 (resolution)	大小 (size)	状态 (status)
`reanalysis/snow/nasa_nesosim-v1.1/`	reanalysis/model	model output (L4)	NESOSIM forced by reanalysis snowfall/winds + PMW drift/SIC	Eulerian 2-layer snow budget model	north	1980 – 2025 (Sep–Apr)	100 km, daily	2.11 GB / 2.11 GB	✅ 完成 (done) 45/45 (100.0%)
`obs/satellite/snow/sysu_snowdepth-pmw-1978-2025/`	retrieval product	L3 (daily gridded)	SMMR, SSM/I, SSMIS	PMW gradient-ratio snow depth with uncertainty	north	1978 – 2025 (Oct–Apr)	25 km polar stereographic, daily	0.88 GB / 0.88 GB	✅ 完成 (done) 1/1 (100.0%)
`reanalysis/snow/eccc_nesosim-mcmc/`	reanalysis/model	model output (L4)	NESOSIM + ERA5/MERRA-2/JRA-55, CloudSat scaling	NESOSIM calibrated by MCMC to snow obs; multi-reanalysis mean + uncertainty	north	1980 – 2023	100 km, daily	4.68 GB / 4.68 GB	✅ 完成 (done) 8/8 (100.0%)
`obs/satellite/snow/sysu_snowdepth-amsr-2002-2024/`	retrieval product	L3 (daily gridded)	AMSR-E, AMSR2	GR(19/7) snow depth with snow-density correction (He et al. 2024)	north	2002 – 2024 (Oct–Apr)	25 km polar stereographic, daily	19.48 GB / 7.65 GB	✅ 完成 (done) 23/23 (100.0%)
`obs/satellite/snow/adasa_amsr2-snowdepth-v1/`	retrieval product	L3 (daily gridded)	AMSR2	Adaptive seasonal algorithm trained on CS-2/IS-2 snow	north	2012 – 2023 (freezing seasons)	daily grid (see files)	5.28 GB / 5.28 GB	✅ 完成 (done) 1/1 (100.0%)
`reanalysis/snow/uw_wassail-v1.0.0/`	reanalysis/model	model output (L4)	WASSAIL forced by ERA5 + satellite drift/SIC	Lagrangian parcel snow-budget model	south	2003-02 – 2025-02	0.75°, daily	16.05 GB / 16.05 GB	✅ 完成 (done) 13/13 (100.0%)
`obs/satellite/snow/legos_laku-snowdepth-v1/`	retrieval product	L3 (monthly gridded)	CryoSat-2 SIRAL (Ku) + ICESat-2 ATLAS	Snow depth = laser minus Ku radar freeboard	north	2018 – 2021 (winter)	25 km, monthly	0.05 GB / 0.05 GB	✅ 完成 (done) 21/21 (100.0%)
`derived-index/snow/awi_snow-clim-w99-amsr2-v1.0/`	derived index/climatology	L4 climatology	AMSR2 + Warren-99 climatology	Merged PMW snow + W99, monthly climatology	north	climatology (Oct–Apr)	25 km EPSG:3411, monthly	0.34 GB / 0.34 GB	✅ 完成 (done) 7/7 (100.0%)

Drift / deformation (漂移/形变)

产品文件夹 (product folder)	类型 (type)	级别 (level)	传感器/平台 (sensor)	方法 (method)	半球 (hemisphere)	时间 (time)	分辨率 (resolution)	大小 (size)	状态 (status)
`obs/satellite/nsidc_nsidc-0116-v4.1/`	retrieval product	L3 (daily/weekly gridded)	AVHRR, SMMR, SSM/I, SSMIS, AMSR-E + IABP buoys + NCEP winds	Max cross-correlation feature tracking, optimal interpolation with buoys & wind	both	1978-11 – latest year	25 km EASE-Grid, daily/weekly	3.08 GB	✅ 完成 (done) 192/192 (100.0%)
`obs/satellite/drift-deformation/mcgill-awi_rgps-deformation-composite-v1.0/`	retrieval product	L3 (3-day gridded)	RADARSAT-1 SAR (RGPS)	Strain rates from RGPS Lagrangian trajectories, composited	north	1997 – 2008 (Jan–Mar)	10 km, 3-day	2.96 GB / 2.96 GB	✅ 完成 (done) 12/12 (100.0%)
`obs/satellite/drift-deformation/eccc_sidrr-asits-v1.3/`	retrieval product	L2/L3 (daily, SAR triangles)	Sentinel-1 SAR, RADARSAT Constellation	ECCC-ASITS SAR motion vectors → triangulated deformation/rotation	north	2017 – 2025-08	SAR triangles, daily	30.16 GB / 13.14 GB	✅ 完成 (done) 9/9 (100.0%)
`obs/satellite/drift-deformation/nsidc_sar-antarctic-velocity-v2/`	retrieval product	L3 (daily gridded)	Sentinel-1 SAR	Feature tracking (nansencenter sea_ice_drift) on S1 pairs	south	2015 – 2023	EASE-Grid South, daily	0.24 GB / 0.24 GB	✅ 完成 (done) 9/9 (100.0%)
`obs/satellite/drift-deformation/utokyo_nh-daily-ice-velocity/`	retrieval product	L3 (daily gridded)	AMSR-E, AMSR2 (36/18 GHz)	Maximum cross-correlation, 60 km template	north	2007 – 2025	60 km, daily	0.21 GB / 0.21 GB	✅ 完成 (done) 6/6 (100.0%)
`obs/satellite/drift-deformation/sysu_amsr2-summer-velocity-v1.0/`	retrieval product	L3 (daily gridded)	AMSR2 36.5 GHz	Continuous max cross-correlation on polarisation difference	north	2013 – 2023 (May–Sep)	daily grid (see files)	0.85 GB / 0.85 GB	✅ 完成 (done) 12/12 (100.0%)
`derived-index/drift-deformation/gsfc_lagrangian-parcel-tracks-v1.0/`	derived index/climatology	L4 (Lagrangian database)	PMW drift/SIC + reanalysis	Daily parcel drift tracks with coincident ice/atmosphere conditions	north	2002 – 2020 (sea-ice years)	daily trajectories	0.00 GB / 130.29 GB	⏳ 进行中 (in progress) 0/74 (0.0%)

Age / type (冰龄/冰型)

产品文件夹 (product folder)	类型 (type)	级别 (level)	传感器/平台 (sensor)	方法 (method)	半球 (hemisphere)	时间 (time)	分辨率 (resolution)	大小 (size)	状态 (status)
`obs/satellite/age-type/nersc_sea-ice-age-cdr-v2.1/`	retrieval product	L4 (monthly gridded)	PMW via OSI SAF SIC + drift CDRs	Lagrangian advection of ice-age fractions with OSI SAF drift/SIC	north	1991 – 2024	monthly grid (see files)	0.14 GB / 0.09 GB	✅ 完成 (done) 34/34 (100.0%)
`obs/satellite/age-type/nsidc_ease-iceage-early-v4.1/`	retrieval product	L4 (weekly gridded)	PMW + buoys (NSIDC-0116 motion)	Lagrangian tracking, annual age classes (spin-up period)	north	1978-10 – 1983	12.5 km EASE-Grid, weekly	0.02 GB / 0.02 GB	✅ 完成 (done) 6/6 (100.0%)
`obs/satellite/age-type/utokyo_nh-daily-sea-ice-age/`	retrieval product	L4 (daily)	AMSR-E, AMSR2	Backward tracking with AMSR velocity + SIC → mean/oldest age	north	2007 – 2022	daily, AMSR grid (text)	4.35 GB / 4.35 GB	✅ 完成 (done) 12/12 (100.0%)
`obs/satellite/age-type/sysu_arctic-ice-type-2002-2021/`	retrieval product	L3 (daily gridded)	FY-3 MWRI + other PMW + scatterometer	K-means clustering → open water / FYI / MYI	north	2002 – 2021	12.5 km, daily	0.06 GB / 0.06 GB	✅ 完成 (done) 1/1 (100.0%)

Melt / albedo / temperature / surface (融化/反照率/温度/表面)

产品文件夹 (product folder)	类型 (type)	级别 (level)	传感器/平台 (sensor)	方法 (method)	半球 (hemisphere)	时间 (time)	分辨率 (resolution)	大小 (size)	状态 (status)
`obs/satellite/melt-albedo-temperature/gsfc_ahra-melt-onset-v5/`	retrieval product	L3 (annual gridded)	SMMR, SSM/I, SSMIS	Advanced Horizontal Range Algorithm on PMW Tb time series	north	1979 – 2022	25 km, annual	0.02 GB / 0.02 GB	✅ 完成 (done) 52/52 (100.0%)
`obs/satellite/melt-albedo-temperature/gapfree-arctic-albedo-v3/`	retrieval product	L4 (daily gap-free)	MODIS Terra + PMW Tb	ML albedo from MODIS (melt-pond aware); gaps from albedo–Tb model	north (>60°N)	2000 – 2024 (May–Sep)	12.5 km polar stereographic, daily	1.17 GB / 1.13 GB	✅ 完成 (done) 25/25 (100.0%)
`obs/satellite/melt-albedo-temperature/dmi_t2m-arctic-seaice-v0.1/`	retrieval product	L4 (monthly gap-free)	AVHRR IST CDR (DMI)	T2m from sea/ice surface temperature CDR via empirical model	north	1982-01 – 2021-05	0.05°, monthly	3.62 GB / 3.62 GB	✅ 完成 (done) 1/1 (100.0%)
`obs/satellite/melt-albedo-temperature/ouc_arctic-freeze-onset/`	retrieval product	L3 (annual gridded)	AMSR-E, AMSR2 + NT2 SIC	Improved PMW freeze-onset algorithm	north	2003 – 2021 (no 2011–12)	12.5 km, annual	0.15 GB / 0.15 GB	✅ 完成 (done) 17/17 (100.0%)
`obs/satellite/melt-albedo-temperature/sysu_ascat-form-drag/`	retrieval product	L3 (daily gridded)	MetOp ASCAT	Surface form drag from scatterometer backscatter (SIC>90%)	north	2006 – 2021 (Nov–Apr)	12.5 km, daily	1.84 GB / 1.84 GB	✅ 完成 (done) 1/1 (100.0%)
`reanalysis/melt-albedo-temperature/ncar_era5-skt-biascorr/`	reanalysis/model	reanalysis, bias-corrected	ERA5 + ML correction	ML correction of winter clear-sky skin-temperature bias	north	1981 – 2018	ERA5 grid, monthly + climatology	1.22 GB / 1.22 GB	✅ 完成 (done) 39/39 (100.0%)
`obs/satellite/melt-albedo-temperature/sysu_quikscat-form-drag/`	retrieval product	L3 (daily gridded)	QuikSCAT SeaWinds	Surface form drag from scatterometer backscatter (SIC>90%)	north	1999 – 2009 (Nov–Apr)	12.5 km, daily	1.12 GB / 1.12 GB	✅ 完成 (done) 1/1 (100.0%)
`obs/satellite/melt-albedo-temperature/whu_antarctic-albedo-viirs-1km-uncertainty-v1/`	retrieval product	L4 (daily uncertainty)	S-NPP VIIRS	Monte Carlo + error propagation uncertainty of the 1 km albedo	south	2012 – 2021	1 km, daily	2.29 GB / 4.66 GB	⏳ 进行中 (in progress) 3/10 (30.0%)
`obs/satellite/melt-albedo-temperature/whu_antarctic-albedo-viirs-1km-v2/`	retrieval product	L4 (daily seamless)	S-NPP VIIRS	MBRI albedo retrieval; cloudy pixels reconstructed	south	2012 – 2021	1 km, daily	0.00 GB / 49.83 GB	⏳ 进行中 (in progress) 0/10 (0.0%)

Leads / polynya / landfast / iceberg (冰间水道/冰间湖/固定冰/冰山)

产品文件夹 (product folder)	类型 (type)	级别 (level)	传感器/平台 (sensor)	方法 (method)	半球 (hemisphere)	时间 (time)	分辨率 (resolution)	大小 (size)	状态 (status)
`obs/satellite/leads-polynya-landfast-iceberg/sysu_deep-aa-polynya-v1.1/`	retrieval product	L4 (daily polynya maps)	AMSR-E/AMSR2 SIC & PSSM	Polynya identification/tracking on SIC thresholds (use SIC60_6.25km_20d)	south	2003 – 2022 (Apr–Oct)	6.25 km, daily	5.20 GB / 5.20 GB	✅ 完成 (done) 29/29 (100.0%)
`obs/satellite/leads-polynya-landfast-iceberg/fio-ouc_arctic-landfast-weekly/`	retrieval product	L3 (weekly gridded, ice-chart analysis)	US NIC ice charts (multi-sensor analyst)	Landfast ice extracted from NIC weekly/biweekly charts	north	1976 – 2026	25 km, weekly	0.02 GB / 0.02 GB	✅ 完成 (done) 4/4 (100.0%)
`obs/satellite/leads-polynya-landfast-iceberg/uwm_viirs-leads-unet/`	retrieval product	L3 (daily lead maps)	S-NPP VIIRS thermal IR	U-Net CNN lead detection	north	2011 – 2021 (Nov–Apr)	~1 km, daily	16.11 GB / 16.11 GB	✅ 完成 (done) 13/13 (100.0%)
`obs/satellite/leads-polynya-landfast-iceberg/sxu-sysu_antarctic-icebergs-s1/`	retrieval product	L3 (annual iceberg vectors)	Sentinel-1 SAR	Incremental random-forest iceberg detection + manual correction	south (<55°S)	2018-10 – 2023-10	vectors ≥0.04 km², annual	0.69 GB / 0.69 GB	✅ 完成 (done) 3/3 (100.0%)
`obs/satellite/leads-polynya-landfast-iceberg/gsfc_is2-lead-fraction-v1.0/`	retrieval product	L2/L3 (along-track + gridded)	ICESat-2 ATLAS (ATL07)	Lead fraction & chord length from ATL07 surface type	both	NH 2018-11–2019-04, SH 2019-05–2019-09	along-track + 25 km	0.46 GB / 0.46 GB	✅ 完成 (done) 14/14 (100.0%)
`obs/satellite/leads-polynya-landfast-iceberg/uwm_modis-leads-unet/`	retrieval product	L3 (daily lead maps)	Terra/Aqua MODIS thermal IR	U-Net CNN lead detection	north	2002 – 2021 (Nov–Apr)	~1 km, daily	0.00 GB / 35.35 GB	⏳ 进行中 (in progress) 0/22 (0.0%)

Brightness temperature (亮温)

产品文件夹 (product folder)	类型 (type)	级别 (level)	传感器/平台 (sensor)	方法 (method)	半球 (hemisphere)	时间 (time)	分辨率 (resolution)	大小 (size)	状态 (status)
`obs/satellite/brightness-temperature/uhh_smap-l1b-tb-arctic/`	obs	L1B → gridded daily Tb	SMAP radiometer (L-band)	SMAP L1B Tb gridded (40° incidence)	north	2015-04 – 2018-05	12.5 km EASE2, daily	0.00 GB / 11.27 GB	⏳ 进行中 (in progress) 0/1150 (0.0%)
`obs/satellite/brightness-temperature/uhh_smos-tb40-arctic-v620/`	obs	L1C → gridded daily Tb	SMOS MIRAS (L-band)	Multi-angle fit to 40° incidence, gridded	north	2015-04 – 2018-05	12.5 km EASE2, daily	0.00 GB / 13.57 GB	⏳ 进行中 (in progress) 0/1154 (0.0%)

Unmodified mirror of public Arctic and Antarctic sea-ice products (satellite observations and retrievals, reanalyses/model products, derived indices and climatologies), collected for validating CICE sea-ice model runs. Files keep their original names; every file is listed in manifest.csv (size, MD5, checksum source, original URL, classification columns, licence, verified on Hugging Face / ModelScope). catalog.csv has one row per product with the full classification and citation.

All credit belongs to the original data producers. Please cite the original products (below) and follow their licences. This mirror is not an official distribution channel; check the source for the latest versions.

Checksums: NSIDC G02202 files are checked against NSIDC .mnf manifests and all Zenodo files against the MD5 published by the Zenodo API. Other providers publish no per-file checksums, so their MD5 was computed on download after an integrity check (NetCDF/HDF opened, gzip/zip tested, TIFF/PNG/PDF magic bytes, no HTML error page). Members unpacked from Zenodo archives carry their own MD5 computed after extraction from an MD5-verified archive.

Classification and folder layout

Every product carries these fields in catalog.csv, manifest.csv and the tables below: type (data_type), category, processing level, sensor/platform, method, hemisphere, time range, resolution, version, DOI, citation, licence.

Type (data_type)

obs: raw or L1/L2 observations (brightness temperatures, along-track L2 retrievals)
retrieval product: L3/L4 gridded satellite geophysical products (concentration, thickness, drift, age, albedo, leads, ...)
reanalysis/model: reanalyses, data-assimilation reconstructions and model-based products (PIOMAS, TOPAZ4-ML, NESOSIM, WASSAIL, Hadley-OI forcing)
derived index/climatology: integrated indices, ensembles, climatologies, ice-edge contours, Lagrangian databases

Category: concentration-extent, thickness-freeboard-volume, snow, drift-deformation, age-type, melt-albedo-temperature (incl. surface roughness/drag), leads-polynya-landfast-iceberg, brightness-temperature.

Folder layout

obs/satellite/<product>/                      products mirrored before batch 4 (kept in place, not moved)
obs/satellite/<category>/<product>/           batch 4 obs + retrieval products (Zenodo)
reanalysis/<product>/                         PIOMAS (before batch 4)
reanalysis/<category>/<product>/              batch 4 reanalysis/model products
derived-index/<product>/                      NSIDC Sea Ice Index (before batch 4)
derived-index/<category>/<product>/           batch 4 indices/climatologies

Product folder names are <producer>_<product>-<version>. Unpacked Zenodo archives keep their members unchanged in a sub-folder named after the archive; archives that could not be usefully unpacked (7z, rar, multi-part zip, nested/mixed contents, shapefiles, text grids) are stored as released.

Product folders

folder	category	type	files uploaded	size (GB)
`obs/satellite/age-type/nersc_sea-ice-age-cdr-v2.1/`	age-type	retrieval product	34/34	0.14
`obs/satellite/age-type/nsidc_ease-iceage-early-v4.1/`	age-type	retrieval product	6/6	0.02
`obs/satellite/age-type/sysu_arctic-ice-type-2002-2021/`	age-type	retrieval product	1/1	0.06
`obs/satellite/age-type/utokyo_nh-daily-sea-ice-age/`	age-type	retrieval product	12/12	4.35
`obs/satellite/brightness-temperature/uhh_smap-l1b-tb-arctic/`	brightness-temperature	obs	0/1150	0.00
`obs/satellite/brightness-temperature/uhh_smos-tb40-arctic-v620/`	brightness-temperature	obs	0/1154	0.00
`derived-index/concentration-extent/esr_antarctic-primary-ice-edge/`	concentration-extent	derived index/climatology	5/5	1.99
`derived-index/concentration-extent/ucla_antarctic-sie-recon-20c/`	concentration-extent	derived index/climatology	6/6	0.41
`derived-index/concentration-extent/uhh_sia-ensemble-osi450a/`	concentration-extent	derived index/climatology	25/25	0.20
`derived-index/concentration-extent/uhh_sie-ensemble-osi450a1/`	concentration-extent	derived index/climatology	25/25	0.17
`derived-index/concentration-extent/uw_arctic-sie-recon-instrumental/`	concentration-extent	derived index/climatology	1/1	0.01
`derived-index/nsidc_g02135-v4/`	concentration-extent	derived index/climatology	65866/65866	25.11
`obs/satellite/concentration-extent/caltech_modis-floe-segmentation-v2/`	concentration-extent	retrieval product	26/26	8.36
`obs/satellite/concentration-extent/dartmouth_is2-gridded-lif-fsd-v6/`	concentration-extent	retrieval product	2/2	2.90
`obs/satellite/concentration-extent/dmi_msc-sic-v0/`	concentration-extent	retrieval product	0/5	0.00
`obs/satellite/concentration-extent/dtu_nimbus6-scams-v1.0/`	concentration-extent	retrieval product	2/2	2.09
`obs/satellite/concentration-extent/snu_landsat8-sic-6.25km/`	concentration-extent	retrieval product	12/12	0.27
`obs/satellite/esacci_sea-ice-concentration/`	concentration-extent	retrieval product	42361/42361	242.91
`obs/satellite/nsidc_g02202-v6/`	concentration-extent	retrieval product	2508/2508	8.85
`obs/satellite/osisaf_osi-430-a/`	concentration-extent	retrieval product	3564/3564	3.87
`obs/satellite/osisaf_osi-430-b/`	concentration-extent	retrieval product	5348/5348	30.31
`obs/satellite/osisaf_osi-450-a1/`	concentration-extent	retrieval product	28439/28439	31.28
`obs/satellite/unibremen_amsr2-asi-6250/`	concentration-extent	retrieval product	10456/10456	18.43
`reanalysis/concentration-extent/ncar_hadley-oi-sst-sic-v6.1/`	concentration-extent	reanalysis/model	3/3	0.95
`reanalysis/concentration-extent/uclouvain_antarctic-recon-1958-2023-v2/`	concentration-extent	reanalysis/model	54/54	0.46
`derived-index/drift-deformation/gsfc_lagrangian-parcel-tracks-v1.0/`	drift-deformation	derived index/climatology	0/74	0.00
`obs/satellite/drift-deformation/eccc_sidrr-asits-v1.3/`	drift-deformation	retrieval product	9/9	30.16
`obs/satellite/drift-deformation/mcgill-awi_rgps-deformation-composite-v1.0/`	drift-deformation	retrieval product	12/12	2.96
`obs/satellite/drift-deformation/nsidc_sar-antarctic-velocity-v2/`	drift-deformation	retrieval product	9/9	0.24
`obs/satellite/drift-deformation/sysu_amsr2-summer-velocity-v1.0/`	drift-deformation	retrieval product	12/12	0.85
`obs/satellite/drift-deformation/utokyo_nh-daily-ice-velocity/`	drift-deformation	retrieval product	6/6	0.21
`obs/satellite/nsidc_nsidc-0116-v4.1/`	drift-deformation	retrieval product	192/192	3.08
`obs/satellite/leads-polynya-landfast-iceberg/fio-ouc_arctic-landfast-weekly/`	leads-polynya-landfast-iceberg	retrieval product	4/4	0.02
`obs/satellite/leads-polynya-landfast-iceberg/gsfc_is2-lead-fraction-v1.0/`	leads-polynya-landfast-iceberg	retrieval product	14/14	0.46
`obs/satellite/leads-polynya-landfast-iceberg/sxu-sysu_antarctic-icebergs-s1/`	leads-polynya-landfast-iceberg	retrieval product	3/3	0.69
`obs/satellite/leads-polynya-landfast-iceberg/sysu_deep-aa-polynya-v1.1/`	leads-polynya-landfast-iceberg	retrieval product	29/29	5.20
`obs/satellite/leads-polynya-landfast-iceberg/uwm_modis-leads-unet/`	leads-polynya-landfast-iceberg	retrieval product	0/22	0.00
`obs/satellite/leads-polynya-landfast-iceberg/uwm_viirs-leads-unet/`	leads-polynya-landfast-iceberg	retrieval product	13/13	16.11
`obs/satellite/melt-albedo-temperature/dmi_t2m-arctic-seaice-v0.1/`	melt-albedo-temperature	retrieval product	1/1	3.62
`obs/satellite/melt-albedo-temperature/gapfree-arctic-albedo-v3/`	melt-albedo-temperature	retrieval product	25/25	1.17
`obs/satellite/melt-albedo-temperature/gsfc_ahra-melt-onset-v5/`	melt-albedo-temperature	retrieval product	52/52	0.02
`obs/satellite/melt-albedo-temperature/ouc_arctic-freeze-onset/`	melt-albedo-temperature	retrieval product	17/17	0.15
`obs/satellite/melt-albedo-temperature/sysu_ascat-form-drag/`	melt-albedo-temperature	retrieval product	1/1	1.84
`obs/satellite/melt-albedo-temperature/sysu_quikscat-form-drag/`	melt-albedo-temperature	retrieval product	1/1	1.12
`obs/satellite/melt-albedo-temperature/whu_antarctic-albedo-viirs-1km-uncertainty-v1/`	melt-albedo-temperature	retrieval product	3/10	2.23
`obs/satellite/melt-albedo-temperature/whu_antarctic-albedo-viirs-1km-v2/`	melt-albedo-temperature	retrieval product	0/10	0.00
`reanalysis/melt-albedo-temperature/ncar_era5-skt-biascorr/`	melt-albedo-temperature	reanalysis/model	39/39	1.22
`derived-index/snow/awi_snow-clim-w99-amsr2-v1.0/`	snow	derived index/climatology	7/7	0.34
`obs/satellite/snow/adasa_amsr2-snowdepth-v1/`	snow	retrieval product	1/1	5.28
`obs/satellite/snow/legos_laku-snowdepth-v1/`	snow	retrieval product	21/21	0.05
`obs/satellite/snow/sysu_snowdepth-amsr-2002-2024/`	snow	retrieval product	23/23	19.48
`obs/satellite/snow/sysu_snowdepth-pmw-1978-2025/`	snow	retrieval product	1/1	0.88
`reanalysis/snow/eccc_nesosim-mcmc/`	snow	reanalysis/model	8/8	4.68
`reanalysis/snow/nasa_nesosim-v1.1/`	snow	reanalysis/model	45/45	2.11
`reanalysis/snow/uw_wassail-v1.0.0/`	snow	reanalysis/model	13/13	16.05
`derived-index/thickness-freeboard-volume/legos_sivol-cdr-v2/`	thickness-freeboard-volume	derived index/climatology	16/16	0.01
`obs/satellite/awi_cs2smos/`	thickness-freeboard-volume	retrieval product	2935/2935	4.61
`obs/satellite/esacci_sea-ice-thickness/`	thickness-freeboard-volume	retrieval product	504/504	0.84
`obs/satellite/jpl_sits-gnssr-l3-antarctic-v1.0/`	thickness-freeboard-volume	retrieval product	134/134	1.27
`obs/satellite/jpl_sits-gnssr-l3-arctic-v1.0/`	thickness-freeboard-volume	retrieval product	307/307	3.67
`obs/satellite/norce-tum_gnssr-freeboard-nh-v1.1/`	thickness-freeboard-volume	retrieval product	7/7	0.00
`obs/satellite/nsidc_atl20-v005/`	thickness-freeboard-volume	retrieval product	186/186	0.58
`obs/satellite/thickness-freeboard-volume/dmi_cs2-tfmra50-sit-2011-2022-v2/`	thickness-freeboard-volume	retrieval product	1/1	0.04
`obs/satellite/thickness-freeboard-volume/gsfc_cs2wfa-antarctic-v0.1.1/`	thickness-freeboard-volume	retrieval product	134/134	0.28
`obs/satellite/thickness-freeboard-volume/jpl_sits-gnssr-l2-antarctic-v1.0/`	thickness-freeboard-volume	obs	8/8	0.55
`obs/satellite/thickness-freeboard-volume/jpl_sits-gnssr-l2-arctic-v1.0/`	thickness-freeboard-volume	obs	8/8	0.68
`obs/satellite/thickness-freeboard-volume/kochi-utokyo_nh-daily-sit-thermo/`	thickness-freeboard-volume	retrieval product	31/31	5.34
`obs/satellite/thickness-freeboard-volume/legos_radar-freeboard-nh-v1.1/`	thickness-freeboard-volume	retrieval product	4/4	2.63
`obs/satellite/thickness-freeboard-volume/legos_sit-cdr-ers-envisat-cs2-v2/`	thickness-freeboard-volume	retrieval product	74/74	6.51
`obs/satellite/thickness-freeboard-volume/legos_sit-cdr-sh-amsrsnow/`	thickness-freeboard-volume	retrieval product	37/37	4.99
`obs/satellite/thickness-freeboard-volume/lerma_pmw-sit-ssmi-v1.0/`	thickness-freeboard-volume	retrieval product	1/1	0.36
`obs/satellite/thickness-freeboard-volume/norce-awi_dasit-75km-octapr/`	thickness-freeboard-volume	retrieval product	20/20	14.51
`obs/satellite/thickness-freeboard-volume/norce_dasit-v100/`	thickness-freeboard-volume	retrieval product	2/2	22.96
`obs/satellite/thickness-freeboard-volume/tju_ice-3d-ml-sit/`	thickness-freeboard-volume	retrieval product	24/24	4.88
`obs/satellite/thickness-freeboard-volume/ucl_daily-pan-arctic-sit/`	thickness-freeboard-volume	retrieval product	33/33	8.12
`obs/satellite/thickness-freeboard-volume/ucl_s3-narrs-v1.0.0/`	thickness-freeboard-volume	obs	2/2	4.06
`obs/satellite/thickness-freeboard-volume/uhh_smos-smap-sit-v1.0/`	thickness-freeboard-volume	retrieval product	577/577	6.20
`obs/satellite/thickness-freeboard-volume/umd_is2sit-summer-v1/`	thickness-freeboard-volume	retrieval product	33/33	0.27
`obs/satellite/thickness-freeboard-volume/umd_is2smgpsit-v1/`	thickness-freeboard-volume	retrieval product	54/54	8.12
`obs/satellite/thickness-freeboard-volume/whu_arctic-sit-5km-v1.0/`	thickness-freeboard-volume	retrieval product	1/1	1.70
`reanalysis/psc_piomas-v2.1/`	thickness-freeboard-volume	reanalysis/model	449/449	2.43
`reanalysis/thickness-freeboard-volume/nersc_topaz4-ml-sit/`	thickness-freeboard-volume	reanalysis/model	0/32	0.00

Products and citations

Concentration / extent (密集度/范围)

`obs/satellite/osisaf_osi-450-a1/` - EUMETSAT OSI SAF Global Sea Ice Concentration CDR v3.1 (OSI-450-a1)

Type / level: retrieval product / L4 (daily gridded, gap-filled CDR)
Sensor / platform: SMMR, SSM/I, SSMIS
Method: Hybrid dynamic-tie-point PMW algorithm + RTM atmospheric correction
Hemisphere / time / resolution: both / 1978-10 – 2020-12 / 25 km EASE2, daily/monthly
Version: v3.1
DOI / reference: 10.15770/EUM_SAF_OSI_0013
Licence: Free and open; attribution to EUMETSAT OSI SAF required (served by MET Norway THREDDS under NLOD and CC BY 4.0).
Citation: OSI SAF (2025): Global Sea Ice Concentration Climate Data Record v3.1, OSI-450-a1, EUMETSAT Ocean and Sea Ice Satellite Application Facility, doi:10.15770/EUM_SAF_OSI_0013. Lavergne, T. et al. (2019), The Cryosphere 13, 49-78, doi:10.5194/tc-13-49-2019.
Coverage in this mirror: NH+SH, EASE2 25 km. Daily files full record 1978-10-25..2020-12-31; monthly and yearly means full record 1978-2020.
Subsetting decision: OSI-450-a1 v3.1 (current release). Full daily record mirrored (backfill of 1978-2004 and 2010-2020 added 2026-10). Continued by OSI-430-a (2021-2025-09).
Source: https://thredds.met.no/thredds/catalog/osisaf/met.no/reprocessed/ice/conc_450a1_files/catalog.html

`obs/satellite/unibremen_amsr2-asi-6250/` - University of Bremen AMSR2 ASI sea ice concentration 6.25 km v5.4

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: AMSR2 (GCOM-W1)
Method: ASI algorithm, 89 GHz polarisation difference
Hemisphere / time / resolution: both / 2012-07 – present / 6.25 km polar stereographic, daily
Version: v5.4
DOI / reference: 10.1594/PANGAEA.898399 (Arctic), 10.1594/PANGAEA.898400 (Antarctic)
Licence: Free for use with citation; PANGAEA editions are CC-BY-4.0.
Citation: Spreen, G., Kaleschke, L., Heygster, G. (2008): Sea ice remote sensing using AMSR-E 89 GHz channels, J. Geophys. Res. 113, C02S03, doi:10.1029/2005JC003384. Melsheimer, C., Spreen, G. (2019): AMSR2 ASI sea ice concentration data, Arctic/Antarctic, version 5.4, PANGAEA.
Coverage in this mirror: NH (n6250) + SH (s6250) polar stereographic 6.25 km, daily NetCDF, 2012-07..present (as of mirror date).
Subsetting decision: AMSR2 starts July 2012 (no 2005-2009 coverage). Full daily record taken because the NetCDF daily files are small (~1.2 MB NH, ~0.6 MB SH; ~9 GB total). Only the netcdf/ subfolders (v5.4) are mirrored; HDF/GeoTIFF/PNG duplicates of the same fields were skipped.
Source: https://data.seaice.uni-bremen.de/amsr2/asi_daygrid_swath/

`derived-index/nsidc_g02135-v4/` - NOAA/NSIDC Sea Ice Index, Version 4 (G02135)

Type / level: derived index/climatology / L4 index (extent/area tables + maps)
Sensor / platform: SMMR, SSM/I, SSMIS
Method: NASA Team SIC → extent (≥15%) / area, daily & monthly, regional
Hemisphere / time / resolution: both / 1978-11 – present / 25 km polar stereographic; tables
Version: v4
DOI / reference: 10.7265/a98x-0f50
Licence: No restrictions on use; please cite.
Citation: Fetterer, F., Knowles, K., Meier, W. N., Savoie, M., Windnagel, A. K., Stafford, T. (2025). Sea Ice Index (G02135, Version 4). NSIDC, Boulder, CO. doi:10.7265/a98x-0f50.
Coverage in this mirror: NH+SH, 1978-present. Daily and monthly extent/area tables (CSV), regional spreadsheets, monthly GeoTIFFs and daily concentration/extent GeoTIFFs (25 km).
Subsetting decision: Full record of data/, seaice_analysis/, monthly and daily GeoTIFFs. Images (PNG) and shapefiles skipped.
Source: https://noaadata.apps.nsidc.org/NOAA/G02135/

`obs/satellite/esacci_sea-ice-concentration/` - ESA Sea Ice CCI High(er)-Resolution Sea Ice Concentration CDR v3.0 (SSM/I & SSMIS, 12.5 km)

Type / level: retrieval product / L4 (daily gridded CDR)
Sensor / platform: SSM/I, SSMIS, AMSR-E, AMSR2
Method: SICCI hybrid algorithm with dynamic tie-points + RTM correction
Hemisphere / time / resolution: both / 1991/2002 – 2020 (SSMI 12.5 km v3.0; AMSR 25/50 km v2.1 2002–2017) / 12.5 / 25 / 50 km EASE2, daily
Version: v3.0 / v2.1
DOI / reference: 10.5285/eade27004395466aaa006135e1b2ad1a
Licence: ESA CCI open data policy (CEDA permitted use: any), citation required.
Citation: Lavergne, T., Sorensen, A., Tonboe, R. T., Kreiner, M., Saldo, R., Birkedal, A., Baordo, F., Aspenes, T., Eastwood, S. (2023): ESA Sea Ice Climate Change Initiative (Sea_Ice_cci): High(er) Resolution Sea Ice Concentration Climate Data Record Version 3 (SSM/I and SSMIS). NERC EDS CEDA. doi:10.5285/eade27004395466aaa006135e1b2ad1a.
Coverage in this mirror: NH+SH daily. SSMI/SSMIS 12.5 km v3.0 full record (L4/ssmi_ssmis/12.5km/v3.0); AMSR-E/AMSR2 25 km and 50 km v2.1, 2002-2017 (L4/amsr/*/v2.1).
Subsetting decision: Full v3.0 record (backfilled 2026-10) and full AMSR v2.1 25/50 km record. Older v2.0 AMSR and phase-1 products skipped (superseded).
Source: https://dap.ceda.ac.uk/neodc/esacci/sea_ice/data/sea_ice_concentration/L4/ssmi_ssmis/12.5km/v3.0/

`obs/satellite/nsidc_g02202-v6/` - NOAA/NSIDC Climate Data Record of Passive Microwave Sea Ice Concentration, Version 6 (G02202)

Type / level: retrieval product / L3 (daily/monthly gridded CDR)
Sensor / platform: SMMR, SSM/I, SSMIS
Method: CDR = max of NASA Team and Bootstrap SIC
Hemisphere / time / resolution: both / 1978-10 – present / 25 km polar stereographic, daily/monthly
Version: v6
DOI / reference: see https://nsidc.org/data/g02202/versions/6 for the Version 6 DOI
Licence: No restrictions on use (NOAA open data); please cite.
Citation: Meier, W. N., Fetterer, F., Windnagel, A. K., Stewart, J. S. (2025). NOAA/NSIDC Climate Data Record of Passive Microwave Sea Ice Concentration, Version 6 (G02202). NSIDC, Boulder, CO.
Coverage in this mirror: NH (psn25) + SH (pss25), 25 km polar stereographic, 1978-10 to present: yearly aggregate files containing all daily fields, plus monthly files, ancillary masks and NSIDC .mnf checksum manifests.
Subsetting decision: Version 6 is the current release (v5 is no longer on the public server). It is served publicly by NOAA@NSIDC without Earthdata login. Full record via the yearly aggregate files (~8.3 GB) + monthly (~0.45 GB); the individual daily files (~15 GB) are exact duplicates of the aggregates and were skipped. Every NetCDF was verified against the NSIDC-provided MD5 (.mnf).
Source: https://noaadata.apps.nsidc.org/NOAA/G02202_V6/

`obs/satellite/osisaf_osi-430-a/` - EUMETSAT OSI SAF Sea Ice Concentration ICDR v3.0 (OSI-430-a)

Type / level: retrieval product / L4 (daily gridded ICDR)
Sensor / platform: SSMIS
Method: OSI-450-a1 algorithm continued (interim CDR)
Hemisphere / time / resolution: both / 2021-01 – 2025-09 / 25 km EASE2, daily/monthly
Version: v3.0
DOI / reference: 10.15770/EUM_SAF_OSI_0014
Licence: Free and open; attribution to EUMETSAT OSI SAF required (served by MET Norway THREDDS under NLOD and CC BY 4.0).
Citation: OSI SAF: Global Sea Ice Concentration Interim Climate Data Record v3.0 (OSI-430-a), EUMETSAT OSI SAF.
Coverage in this mirror: NH+SH, EASE2 25 km, daily 2021-01-01..2025-09-30 (production stopped 2025-09-30) plus monthly/yearly means.
Subsetting decision: Full record mirrored. Extends OSI-450-a1 beyond 2020.
Source: https://thredds.met.no/thredds/catalog/osisaf/met.no/reprocessed/ice/conc_cra_files/catalog.html

`obs/satellite/osisaf_osi-430-b/` - EUMETSAT OSI SAF Sea Ice Concentration ICDR v2.0 (OSI-430-b, deprecated)

Type / level: retrieval product / L4 (daily gridded ICDR, deprecated)
Sensor / platform: SSMIS
Method: OSI-450 v2 algorithm continued (interim CDR)
Hemisphere / time / resolution: both / 2016 – 2023-05 / 25 km EASE2, daily
Version: v2.0
DOI / reference: 10.15770/EUM_SAF_OSI_NRT_2008
Licence: Free and open; attribution to EUMETSAT OSI SAF required (served by MET Norway THREDDS under NLOD and CC BY 4.0).
Citation: OSI SAF: Global Sea Ice Concentration Interim Climate Data Record v2.0 (OSI-430-b), EUMETSAT OSI SAF.
Coverage in this mirror: NH+SH, EASE2 25 km, daily 2016..2023-05.
Subsetting decision: Deprecated (superseded by OSI-450-a1/OSI-430-a); full record mirrored on request for completeness.
Source: https://thredds.met.no/thredds/catalog/osisaf/met.no/reprocessed/ice/conc_crb_files/catalog.html

`obs/satellite/concentration-extent/dtu_nimbus6-scams-v1.0/` - Sea Ice dataset for the NIMBUS-6 Scanning Microwave Spectrometer (SCAMS) data from 1975 and 1976, version 1.0 (v1.0)

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: Nimbus-6 SCAMS (22.2/31.7 GHz)
Method: SIC and ice type from 22.235/31.65 GHz Tb
Hemisphere / time / resolution: both / 1975-06 – 1976-06 / daily grid (see files)
Version: v1.0
DOI / reference: 10.5281/zenodo.15109768 (version; concept 10.5281/zenodo.15109767)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Kolbe, Wiebke Margitta; Tonboe, Rasmus Tage (2025): Sea Ice dataset for the NIMBUS-6 Scanning Microwave Spectrometer (SCAMS) data from 1975 and 1976, version 1.0 (Version v1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.15109768
Coverage in this mirror: both; 1975-06 – 1976-06; daily grid (see files). Zenodo record(s) total 2 files, 2.09 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/15109768

`reanalysis/concentration-extent/ncar_hadley-oi-sst-sic-v6.1/` - Merged Hadley-OI sea surface temperature and sea ice concentration data set (6.1)

Type / level: reanalysis/model / L4 analysis (model forcing)
Sensor / platform: HadISST1 (in situ + satellite) + NOAA OI.v2
Method: Merged HadISST1/OI.v2 SST & SIC, anomaly-based blending (CAM AMIP forcing)
Hemisphere / time / resolution: both (global) / 1870-01 – 2022-06 / 1°, monthly
Version: v6.1
DOI / reference: 10.5281/zenodo.15501326 (version; concept 10.5281/zenodo.15501325)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Shea, Dennis; Hurrell, Jim; Phillips, Adam (2024): Merged Hadley-OI sea surface temperature and sea ice concentration data set (Version 6.1) [Data set]. Zenodo. https://doi.org/10.5065/r33v-sv91
Coverage in this mirror: both (global); 1870-01 – 2022-06; 1°, monthly. Zenodo record(s) total 3 files, 0.95 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/15501326

`derived-index/concentration-extent/uhh_sie-ensemble-osi450a1/` - Sea Ice Extent Ensemble for Uncertainty Quantification

Type / level: derived index/climatology / L4 index (extent ensemble)
Sensor / platform: SMMR, SSM/I, SSMIS (OSI-450-a1)
Method: SIE (SIC≥15%) with 50-member correlated-noise ensemble
Hemisphere / time / resolution: both (+regions) / 1978 – 2020 / daily regional totals (CSV)
Version: 2026-08 release
DOI / reference: 10.5281/zenodo.21793108 (version; concept 10.5281/zenodo.21793107)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Wernecke, Andreas (2026): Sea Ice Extent Ensemble for Uncertainty Quantification [Data set]. Zenodo. https://doi.org/10.5281/zenodo.21793108
Coverage in this mirror: both (+regions); 1978 – 2020; daily regional totals (CSV). Zenodo record(s) total 25 files, 0.17 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/21793108

`derived-index/concentration-extent/uhh_sia-ensemble-osi450a/` - Sea Ice Area Ensemble for Uncertainty Quantification

Type / level: derived index/climatology / L4 index (area ensemble)
Sensor / platform: SMMR, SSM/I, SSMIS (OSI-450a)
Method: SIA with 50-member correlated-noise ensemble
Hemisphere / time / resolution: both (+regions) / 1978 – 2020 / daily regional totals (CSV)
Version: 2025-10 release
DOI / reference: 10.5281/zenodo.17464351 (version; concept 10.5281/zenodo.17464350)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Wernecke, Andreas (2025): Sea Ice Area Ensemble for Uncertainty Quantification [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17464351
Coverage in this mirror: both (+regions); 1978 – 2020; daily regional totals (CSV). Zenodo record(s) total 25 files, 0.20 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17464351

`obs/satellite/concentration-extent/dartmouth_is2-gridded-lif-fsd-v6/` - Gridded sea ice concentration, floe size distribution, and wave data from the ICESat-2 altimeter: October 2018-December 2024.

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: ICESat-2 ATLAS
Method: Along-track ice/water class → linear ice fraction, floe size, waves
Hemisphere / time / resolution: both / 2018-10 – 2024-12 / 25 km, monthly
Version: v6 files
DOI / reference: 10.5281/zenodo.17305285 (version; concept 10.5281/zenodo.17305284)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Horvat, Christopher; Buckley, Ellen (2025): Gridded sea ice concentration, floe size distribution, and wave data from the ICESat-2 altimeter: October 2018-December 2024. [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17305285
Coverage in this mirror: both; 2018-10 – 2024-12; 25 km, monthly. Zenodo record(s) total 2 files, 2.90 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17305285

`derived-index/concentration-extent/esr_antarctic-primary-ice-edge/` - Primary Sea Ice Edge from Satellite Passive Microwave Observations

Type / level: derived index/climatology / L4 index (ice-edge contours)
Sensor / platform: SSM/I, AMSR-E, AMSR2
Method: Northernmost 15% SIC contour (NT2/ARTIST), smoothed/unsmoothed
Hemisphere / time / resolution: south / SSM/I – AMSR2 eras (see files) / daily contours (MATLAB)
Version: 2023-11 release
DOI / reference: 10.5281/zenodo.10140318 (version; concept 10.5281/zenodo.3669405)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Springer, Scott; Padman, Laurence; Howard, Susan (2023): Primary Sea Ice Edge from Satellite Passive Microwave Observations [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10140318
Coverage in this mirror: south; SSM/I – AMSR2 eras (see files); daily contours (MATLAB). Zenodo record(s) total 5 files, 1.99 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/10140318

`obs/satellite/concentration-extent/snu_landsat8-sic-6.25km/` - Arctic sea ice concentration data record in 6.25 km polar stereographic grid from three-year Landsat-8 imagery

Type / level: retrieval product / L3 (gridded scenes)
Sensor / platform: Landsat-8 OLI
Method: 30 m ice/water classification aggregated to 6.25 km SIC
Hemisphere / time / resolution: north (12 regions) / three years of Landsat-8 (see files) / 6.25 km polar stereographic
Version: 2024-04 release
DOI / reference: 10.5281/zenodo.12750969 (version; concept 10.5281/zenodo.10973297)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Jung, Hee-Sung; Lee, Sang-Moo; Kim, Joo-Hong; Lee, Kyungsoo (2024): Arctic sea ice concentration data record in 6.25 km polar stereographic grid from three-year Landsat-8 imagery [Data set]. Zenodo. https://doi.org/10.5281/zenodo.12750969
Coverage in this mirror: north (12 regions); three years of Landsat-8 (see files); 6.25 km polar stereographic. Zenodo record(s) total 12 files, 0.27 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/12750969

`obs/satellite/concentration-extent/caltech_modis-floe-segmentation-v2/` - Ice Floe Segmentation of MODIS imagery version 2 (2)

Type / level: retrieval product / L3 (daily floe maps)
Sensor / platform: MODIS Terra true colour
Method: Floe segmentation → floe properties, SIC, floe-size distribution
Hemisphere / time / resolution: north (Beaufort Sea) / 2000 – 2022 (Mar–Sep) / 256 m, daily
Version: v2
DOI / reference: 10.5281/zenodo.22885838 (version; concept 10.5281/zenodo.11553699)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Buckley, Ellen; Wilhelmus, Monica (2026): Ice Floe Segmentation of MODIS imagery version 2 (Version 2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22885838
Coverage in this mirror: north (Beaufort Sea); 2000 – 2022 (Mar–Sep); 256 m, daily. Zenodo record(s) total 26 files, 2.17 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/22885838

`reanalysis/concentration-extent/uclouvain_antarctic-recon-1958-2023-v2/` - Antarctic sea ice reconstructions over 1958-2023 (Version 2 includes the ensemble mean)

Type / level: reanalysis/model / reconstruction (data assimilation)
Sensor / platform: 9 ESM priors + station/ice-core data
Method: Data-assimilation reconstruction, annual anomalies + uncertainty
Hemisphere / time / resolution: south / 1958 – 2023 / annual fields
Version: Version 2
DOI / reference: 10.5281/zenodo.13861997 (version; concept 10.5281/zenodo.13828874)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Dalaiden, Quentin; Goosse, Hugues (2024): Antarctic sea ice reconstructions over 1958-2023 (Version Version 2 includes the ensemble mean) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.13861997
Coverage in this mirror: south; 1958 – 2023; annual fields. Zenodo record(s) total 54 files, 0.46 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/13861997

`derived-index/concentration-extent/ucla_antarctic-sie-recon-20c/` - 20th Century Antarctic sea ice extent anomaly reconstruction by sector (1.0)

Type / level: derived index/climatology / L4 index (reconstruction)
Sensor / platform: statistical reconstruction
Method: Sector SIE anomaly reconstruction, 2500 members
Hemisphere / time / resolution: south (sectors) / 20th century / monthly (CSV)
Version: v1.0
DOI / reference: 10.5281/zenodo.7971734 (version; concept 10.5281/zenodo.7971733)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Thomas Maierhofer; Raphael, Marilyn; Handcock, Mark (2023): 20th Century Antarctic sea ice extent anomaly reconstruction by sector (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.7971734
Coverage in this mirror: south (sectors); 20th century; monthly (CSV). Zenodo record(s) total 6 files, 0.41 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/7971734

`derived-index/concentration-extent/uw_arctic-sie-recon-instrumental/` - Total Arctic sea ice extent during the Instrumental Era (1)

Type / level: derived index/climatology / L4 index (reconstruction)
Sensor / platform: CCSM4/MPI priors + HadCRUT/BE/GISTEMP
Method: Data-assimilation reconstructions of total Arctic SIE
Hemisphere / time / resolution: north / instrumental era (see file) / annual totals
Version: v1
DOI / reference: 10.5281/zenodo.3717240 (version; concept 10.5281/zenodo.3717239)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Brennan, M. Kathleen; Hakim, Gregory J.; Blanchard-Wrigglesworth, Edward (2020): Total Arctic sea ice extent during the Instrumental Era (Version 1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.3717240
Coverage in this mirror: north; instrumental era (see file); annual totals. Zenodo record(s) total 1 files, 0.01 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/3717240

`obs/satellite/concentration-extent/dmi_msc-sic-v0/` - Danish Meteorological Institute Multisource Sea Ice Concentration Composite (DMI-MSC-SIC) (v0)

Type / level: retrieval product / L4 (daily gap-free composite)
Sensor / platform: SMMR/SSMI/SSMIS, AMSR-E/AMSR2 + NIC/FMI/SMHI ice charts
Method: Multi-source composite (OSI-450-a, SICCI-HR, OSI-458, charts), bias-corrected, gap-filled
Hemisphere / time / resolution: both (global) / 1982-01 – 2024-12 / 0.05° lat/lon, daily
Version: v0
DOI / reference: 10.5281/zenodo.17100178 (version; concept 10.5281/zenodo.17100177)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Olsen, Ida Lundtorp; Nielsen-Englyst, Pia (2025): Danish Meteorological Institute Multisource Sea Ice Concentration Composite (DMI-MSC-SIC) (Version v0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17100178
Coverage in this mirror: both (global); 1982-01 – 2024-12; 0.05° lat/lon, daily. Zenodo record(s) total 5 files, 141.30 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17100178

Thickness / freeboard / volume (厚度/干舷/体积)

`obs/satellite/awi_cs2smos/` - AWI CryoSat-2/SMOS merged sea ice thickness (CS2SMOS) v2.06

Type / level: retrieval product / L4 (weekly merged)
Sensor / platform: CryoSat-2 SIRAL + SMOS MIRAS
Method: Optimal interpolation of CS-2 radar-freeboard SIT and SMOS thin-ice SIT
Hemisphere / time / resolution: north / 2010-11 – 2025-04 (Oct–Apr) / 25 km EASE2, weekly
Version: v2.06
DOI / reference: (no dataset DOI; reference Ricker et al. 2017, doi:10.5194/tc-11-1607-2017)
Licence: Free and anonymous access; acknowledgement required: 'The production of the merged CryoSat-SMOS sea ice thickness data was funded by the ESA project SMOS & CryoSat-2 Sea Ice Data Product Processing and Dissemination Service, and data from DATE to DATE were obtained from AWI.'
Citation: Ricker, R., Hendricks, S., Kaleschke, L., Tian-Kunze, X., King, J., Haas, C. (2017): A weekly Arctic sea-ice thickness data record from merged CryoSat-2 and SMOS satellite data, The Cryosphere 11, 1607-1623, doi:10.5194/tc-11-1607-2017.
Coverage in this mirror: NH only (Arctic), weekly 25 km EASE2, winter months (Oct-Apr) from 2010/11 to 2024/25; 'r' = reprocessed, 'o' = operational files.
Subsetting decision: CS2SMOS starts in late 2010 (CryoSat-2 + SMOS), so it does NOT cover the 2005-2009 CICE test period. Full v206 record mirrored (small). The newer v300 (CS3SMOS, from 2025/26) was not found on the portal at mirror time.
Source: https://data.seaiceportal.de/data/cs2smos_awi/v206/

`reanalysis/psc_piomas-v2.1/` - PIOMAS (Pan-Arctic Ice Ocean Modeling and Assimilation System) v2.1 thickness and volume

Type / level: reanalysis/model / model reanalysis
Sensor / platform: PIOMAS ice-ocean model (assimilates SIC, SST)
Method: Coupled ice-ocean model with SIC/SST assimilation, NCEP forcing
Hemisphere / time / resolution: north / 1978 – present / PIOMAS grid 360×120, daily/monthly
Version: v2.1
DOI / reference: (no dataset DOI; references below)
Licence: Freely provided by the Polar Science Center, APL/UW; please cite. Note PIOMAS is a model reanalysis, not a direct observation.
Citation: Zhang, J., Rothrock, D. A. (2003): Modeling global sea ice with a thickness and enthalpy distribution model in generalized curvilinear coordinates, Mon. Wea. Rev. 131, 845-861. Schweiger, A. et al. (2011): Uncertainty in modeled Arctic sea ice volume, J. Geophys. Res. 116, C00D06, doi:10.1029/2011JC007084.
Coverage in this mirror: Arctic, PIOMAS model grid (360x120), 1978-present. heff = monthly mean effective thickness; hiday = daily thickness; area = monthly/daily concentration; plus daily/monthly volume and mean-thickness time series and grid/reader utilities.
Subsetting decision: Full record of the thickness (heff, hiday), area and volume products. Other PIOMAS fields (ocean T/S/velocity, snow, ice velocity, forcing) were skipped as out of scope.
Source: https://pscfiles.apl.washington.edu/zhang/PIOMAS/data/v2.1/ and https://psc.apl.uw.edu/research/projects/arctic-sea-ice-volume-anomaly/data/

`obs/satellite/esacci_sea-ice-thickness/` - ESA Sea Ice CCI sea ice thickness L3C v4.0 (ERS-2, Envisat, CryoSat-2)

Type / level: retrieval product / L3C (monthly gridded)
Sensor / platform: ERS-2 RA, Envisat RA-2, CryoSat-2 SIRAL
Method: Radar freeboard + W99m/AMSR snow, hydrostatic equilibrium
Hemisphere / time / resolution: both / ERS-2 1990s–2003, Envisat 2002–2012, CS-2 2010–2024 / 25 km EASE2 (NH) / 50 km (SH), monthly
Version: v4.0
DOI / reference: CEDA records: Envisat NH 10.5285/a0120f31bcd94f8da74ddfebe658773d, CryoSat-2 NH 10.5285/887ebbe3aa4345f3bbcbe2dd1834d955 (see catalogue.ceda.ac.uk for SH/ERS-2 records)
Licence: ESA CCI open data policy (CEDA permitted use: any), citation required.
Citation: Hendricks, S., Paul, S., Rinne, E. (2025): ESA Sea Ice Climate Change Initiative (Sea_Ice_cci): sea ice thickness from Envisat / CryoSat-2 / ERS-2 on a monthly grid (L3C), v4.0. NERC EDS CEDA.
Coverage in this mirror: Monthly L3C grids. Envisat NH+SH 2002-2012, CryoSat-2 NH+SH 2010-2024, ERS-2 NH 1990s-2003. Covers 2005-2009 via Envisat.
Subsetting decision: Only the latest version v4.0 mirrored (full record, <1 GB). Older v2.0/v3.0, L2P along-track and the drift-aware thickness product were skipped.
Source: https://dap.ceda.ac.uk/neodc/esacci/sea_ice/data/sea_ice_thickness/L3C/

`obs/satellite/nsidc_atl20-v005/` - ICESat-2 ATL20 Daily and Monthly Gridded Sea Ice Freeboard, Version 5

Type / level: retrieval product / L3B (gridded)
Sensor / platform: ICESat-2 ATLAS
Method: ATL10 laser total freeboard gridded daily/monthly
Hemisphere / time / resolution: both / 2018-10 – present / 25 km polar stereographic, daily/monthly
Version: v005
DOI / reference: 10.5067/ATLAS/ATL20.005
Licence: NASA open data policy, no restrictions; please cite.
Citation: Petty, A. A., Kurtz, N. T., Kwok, R., Markus, T., Neumann, T. A., Keeney, N. (2023). ATLAS/ICESat-2 L3B Daily and Monthly Gridded Sea Ice Freeboard, Version 5 (ATL20). NASA NSIDC DAAC. doi:10.5067/ATLAS/ATL20.005.
Coverage in this mirror: NH (ATL20-01) + SH (ATL20-02), 25 km, monthly granules (with daily fields inside), 2018-10 to present.
Subsetting decision: ICESat-2 starts in Oct 2018 (no 2005-2009 coverage). Full record taken (gridded product, small). ATL10 along-track freeboard was skipped (very large); ATL20 is the gridded product.
Source: https://data.nsidc.earthdatacloud.nasa.gov/nsidc-cumulus-prod-protected/ATLAS/ATL20/005/ (Earthdata login)

`obs/satellite/jpl_sits-gnssr-l3-antarctic-v1.0/` - SITS Level-3 Antarctic GNSS-R Sea-Ice Thickness Products, Version 1.0 (JPL / Univ. of Hawaii, Spire GNSS-R)

Type / level: retrieval product / L3 (7-day gridded)
Sensor / platform: Spire GNSS-R (grazing + near-nadir)
Method: GNSS-R reflectivity → thickness, multi-observable combination
Hemisphere / time / resolution: south / 2022-10 – 2024-01 / 12.5 km EASE2 South, 7-day
Version: v1.0
DOI / reference: 10.5281/zenodo.23248278 (version), concept DOI 10.5281/zenodo.23248277
Licence: CC BY 4.0 (Zenodo record licence). Note: the producers' README says the final rights statement was to be approved separately with attention to commercial Spire source data and NASA CSDA data-use terms; the published Zenodo licence is CC BY 4.0.
Citation: Rodriguez-Alvarez, N., Kim, S. (2026): SITS Level-3 Antarctic GNSS-R Sea-Ice Thickness Products, Version 1.0 [Data set], Zenodo, doi:10.5281/zenodo.23248278. Method: Kim, S., Rodriguez-Alvarez, N., Oudrhiri, K. (2026): Sea-Ice Thickness Retrieval Using Integrated Grazing-Angle and Near-Nadir GNSS-R Reflectivity From the Spire Constellation, IEEE Trans. Geosci. Remote Sens. 64, 5802117, doi:10.1109/TGRS.2026.3714677. Funded by NASA CSDA (ROSES NNH22ZDA001N-A.44, Task Order 80NM0018F0617).
Coverage in this mirror: SH only, EASE-Grid 2.0 South (EPSG:6932) 560x560, 12.5 km. 65 consecutive non-overlapping 7-day windows 2022-10-15 to 2024-01-12, in two families: combined (multi-observable SIT, salinity, relative density, availability/used masks) and single (SIT/salinity/density per GNSS-R type RCS, NRCS, GRZ L1, GRZ L2). 130 NetCDF + 4 docs, 1.27 GB.
Subsetting decision: Full record mirrored. The two Zenodo ZIPs (product_L3_Antarctic_combined_v100.zip, product_L3_Antarctic_single_v100.zip) were MD5-checked against Zenodo, then unpacked: the NetCDF members are stored unchanged (original names, zip CRC verified) under product_L3_Antarctic_combined/ and product_L3_Antarctic_single/. Producer README / processing description / data dictionary / file inventory are included. No SIT uncertainty or reflectivity fields in the Antarctic release. Packing caveat: science variables are int32 with scale_factor=0.001 in the files, but decoded values are 1000x too large (e.g. SIT up to ~4800 'm'); divide decoded SIT/salinity/density by a further 1000 (effective scale 1e-6; SIT 0-5 m, relative density ~0.84-1.2). Spire_L1B_available/used are int8 bit masks (1=RCS,2=NRCS,4=GRZ L1,8=GRZ L2) with _FillValue=0 that is also the valid zero state. Grid CRS is not in the files (use the EPSG code above); filename token '12km' means 12.5 km.
Source: https://zenodo.org/records/23248278

`obs/satellite/jpl_sits-gnssr-l3-arctic-v1.0/` - SITS Level-3 Arctic GNSS-R Sea-Ice Thickness Products, Version 1.0 (JPL / Univ. of Hawaii, Spire GNSS-R)

Type / level: retrieval product / L3 (7-day gridded)
Sensor / platform: Spire GNSS-R (grazing + near-nadir)
Method: GNSS-R reflectivity → thickness, multi-observable combination
Hemisphere / time / resolution: north / 2022-10 – 2024-01 (winters) / 12.5 km EASE2 North, 7-day
Version: v1.0
DOI / reference: 10.5281/zenodo.23248136 (version), concept DOI 10.5281/zenodo.23248135
Licence: CC BY 4.0 (Zenodo record licence). Note: the producers' README says the final rights statement was to be approved separately with attention to commercial Spire source data and NASA CSDA data-use terms; the published Zenodo licence is CC BY 4.0.
Citation: Rodriguez-Alvarez, N., Kim, S. (2026): SITS Level-3 Arctic GNSS-R Sea-Ice Thickness Products, Version 1.0 [Data set], Zenodo, doi:10.5281/zenodo.23248136. Method: Kim, S., Rodriguez-Alvarez, N., Oudrhiri, K. (2026): Sea-Ice Thickness Retrieval Using Integrated Grazing-Angle and Near-Nadir GNSS-R Reflectivity From the Spire Constellation, IEEE Trans. Geosci. Remote Sens. 64, 5802117, doi:10.1109/TGRS.2026.3714677. Funded by NASA CSDA (ROSES NNH22ZDA001N-A.44, Task Order 80NM0018F0617).
Coverage in this mirror: NH only, EASE-Grid 2.0 North (EPSG:6931) 720x720, 12.5 km, two winter seasons. combined: 266 rolling 7-day windows (start advances 1 day) 2022-10-15..2023-04-15 and 2023-10-15..2024-01-17 (SIT, salinity, relative density, masks; SIT uncertainty and reflectivity mean+SEM only in 36 files). single: 37 non-overlapping 7-day windows 2022-10-15..2023-04-14 and 2023-10-21..2024-01-05 (per-type SIT, salinity, density). 303 NetCDF + 4 docs, 3.67 GB.
Subsetting decision: Full record mirrored (companion of the Antarctic L3 record; the 681 MB Level-2 along-track Arctic record 10.5281/zenodo.21937449 and the 547 MB Antarctic L2 record 10.5281/zenodo.23248020 were not mirrored). Zenodo ZIPs MD5-checked then unpacked; NetCDF members stored unchanged under product_L3_Arctic_combined/ and product_L3_Arctic_single/. Note: single file 20230408_20230414 is only 50 KB as released (nearly empty), Variable sets differ between files as released: only 36 of the 266 combined files carry sea_ice_thickness_uncertainty and the reflectivity mean/SEM fields (GRZ L2 reflectivity is all-fill there); the other 230 combined files have SIT, salinity, density and masks only. In all 37 single files sea_ice_thickness_uncertainty is present but all-fill. Packing caveat: science variables are int32 with scale_factor=0.001 in the files, but decoded values are 1000x too large (e.g. SIT up to ~4800 'm'); divide decoded SIT/salinity/density by a further 1000 (effective scale 1e-6; SIT 0-5 m, relative density ~0.84-1.2). Spire_L1B_available/used are int8 bit masks (1=RCS,2=NRCS,4=GRZ L1,8=GRZ L2) with _FillValue=0 that is also the valid zero state. Grid CRS is not in the files (use the EPSG code above); filename token '12km' means 12.5 km.
Source: https://zenodo.org/records/23248136

`obs/satellite/norce-tum_gnssr-freeboard-nh-v1.1/` - GNSS-R Monthly Gridded Arctic Sea Ice Freeboard 2023-2024, v1.1 (NORCE / TU Munich, Spire grazing-angle GNSS-R, ESA EDAP+)

Type / level: retrieval product / L3C (monthly gridded)
Sensor / platform: Spire grazing-angle GNSS-R
Method: GNSS-R phase altimetry SSH → freeboard (experimental)
Hemisphere / time / resolution: north / 2023-10 – 2024-04 / 50 km EASE2 North, monthly
Version: v1.1
DOI / reference: 10.5281/zenodo.22802123 (v1.1), concept DOI 10.5281/zenodo.20809116
Licence: CC BY 4.0 (Zenodo record licence).
Citation: Ricker, R., Mueller, F. L. (2026): GNSS-R Monthly Gridded Arctic Sea Ice Freeboard 2023-2024 (Version 1.1) [Data set], Zenodo, doi:10.5281/zenodo.22802123. Associated paper: Ricker, R., Mueller, F. L., Talpe, M., Cartwright, J., Halsall, K., Hansen, E., Seitz, F., De Laurentiis, L., Di Bella, A.: Exploring high-resolution grazing-angle GNSS-R for retrieval of basin-scale sea ice freeboard, preprint, SSRN, doi:10.2139/ssrn.6991955.
Coverage in this mirror: NH only (Arctic), EASE-Grid 2.0 North (EPSG:6931, CF grid_mapping included) 216x216, 50 km; ice cells reach ~55N. Monthly L3C fields Oct 2023 to Apr 2024 (7 files, one winter season). Variables: gnssr_freeboard, gnssr_freeboard_smoothed, gnssr_freeboard_std, gnssr_freeboard_cnt, gnssr_freeboard_uncertainty (m), sea_ice_concentration, sea_surface_height, lead_fraction, lat/lon, time_bnds. 7 NetCDF, 4.28 MB.
Subsetting decision: Full record mirrored (latest version 1.1; superseded v1.0 record 10.5281/zenodo.20809117 of 2026-06-23 not mirrored). Files uploaded unchanged (plain NetCDF, no archive). Experimental product. Caveats: sea_ice_concentration is labelled units "1" but holds percent (70-100, cells below 70% are masked); the global attribute product_version reads "v1p0_rev" although the files and record are v1.1; freeboard has a few small negative values (16-46 cells per month, down to about -0.1 m) as released; missing values are NaN. Not from the JPL/UH SITS group (see jpl_sits-gnssr-l3-* folders for their thickness products).
Source: https://zenodo.org/records/22802123

`obs/satellite/thickness-freeboard-volume/legos_sit-cdr-ers-envisat-cs2-v2/` - Arctic and Antarctic sea ice thickness climate data record from ERS-1, ERS-2, Envisat and CryoSat-2 (v2)

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: ERS-1/2 RA, Envisat RA-2, CryoSat-2 SIRAL
Method: Radar freeboard (CS-2 calibrated) + snow options, hydrostatic
Hemisphere / time / resolution: both / 1993/94 – 2023 / monthly grid (see files)
Version: v2
DOI / reference: 10.5281/zenodo.12783561 (version; concept 10.5281/zenodo.12783560)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Bocquet, Marion; Centre National d'Études Spatiales; Laboratoire d'Études en Géophysique et Océanographie Spatiales (2024): Arctic and Antarctic sea ice thickness climate data record from ERS-1, ERS-2, Envisat and CryoSat-2 (Version v2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.12783561
Coverage in this mirror: both; 1993/94 – 2023; monthly grid (see files). Zenodo record(s) total 74 files, 6.20 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/12783561

`derived-index/thickness-freeboard-volume/legos_sivol-cdr-v2/` - Arctic and Antarctic sea ice volume climate data record from ERS-1, ERS-2, Envisat and CryoSat-2 (v2)

Type / level: derived index/climatology / L4 index (integrated volume)
Sensor / platform: ERS-1/2, Envisat, CryoSat-2
Method: Integrated sea ice volume from the LEGOS SIT CDR, regional, with uncertainty
Hemisphere / time / resolution: both / 1993 – 2023 / monthly regional totals (CSV)
Version: v2
DOI / reference: 10.5281/zenodo.20525445 (version; concept 10.5281/zenodo.20525444)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Bocquet, Marion; Centre National d'Études Spatiales; Laboratoire d'Études en Géophysique et Océanographie Spatiales; Fleury, Sara; Rémy, Frédérique; Piras, Fanny; Boulard, Camille (2026): Arctic and Antarctic sea ice volume climate data record from ERS-1, ERS-2, Envisat and CryoSat-2 (Version v2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.20525445
Coverage in this mirror: both; 1993 – 2023; monthly regional totals (CSV). Zenodo record(s) total 16 files, 0.01 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/20525445

`obs/satellite/thickness-freeboard-volume/legos_radar-freeboard-nh-v1.1/` - Arctic sea ice radar freeboard from ERS-1, ERS-2, Envisat and CryoSat-2 (v1.1)

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: ERS-1/2 RA, Envisat RA-2, CryoSat-2 SIRAL
Method: Radar freeboard; ERS/Envisat calibrated to CS-2 SAR TFMRA50
Hemisphere / time / resolution: north / 1993 – 2021 / monthly grid (see files)
Version: v1.1
DOI / reference: 10.5281/zenodo.8063431 (version; concept 10.5281/zenodo.7712502)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Bocquet, Marion; Fleury, Sara (2023): Arctic sea ice radar freeboard from ERS-1, ERS-2, Envisat and CryoSat-2 (Version v1.1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.8063431
Coverage in this mirror: north; 1993 – 2021; monthly grid (see files). Zenodo record(s) total 4 files, 2.63 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/8063431

`obs/satellite/thickness-freeboard-volume/legos_sit-cdr-sh-amsrsnow/` - Antarctic sea ice thickness climate data record (ERS-1, ERS-2, Envisat, CryoSat-2) using a snow depth climatology

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: ERS-1/2 RA, Envisat RA-2, CryoSat-2 SIRAL
Method: Radar freeboard + SI-CCI AMSR snow climatology, hydrostatic
Hemisphere / time / resolution: south / 1994-04 – 2023-06 / monthly grid (see files)
Version: 2024-02 release
DOI / reference: 10.5281/zenodo.10610697 (version; concept 10.5281/zenodo.10610696)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Bocquet, Marion; Fleury, Sara (2024): Antarctic sea ice thickness climate data record (ERS-1, ERS-2, Envisat, CryoSat-2) using a snow depth climatology [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10610697
Coverage in this mirror: south; 1994-04 – 2023-06; monthly grid (see files). Zenodo record(s) total 37 files, 0.44 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/10610697

`obs/satellite/thickness-freeboard-volume/gsfc_cs2wfa-antarctic-v0.1.1/` - Antarctic Sea Ice Thickness Estimates from CryoSat-2: 2010-2021 (0.1.1)

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: CryoSat-2 SIRAL SAR/SARIn (Baseline-D)
Method: Waveform fitting (CS2WFA) snow & ice freeboard, hydrostatic
Hemisphere / time / resolution: south / 2010-07 – 2021-08 / 25 km polar stereographic, monthly
Version: v0.1.1 (preliminary)
DOI / reference: 10.5281/zenodo.7327711 (version; concept 10.5281/zenodo.7327710)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Fons, Steven; Kurtz, Nathan; Bagnardi, Marco (2022): Antarctic Sea Ice Thickness Estimates from CryoSat-2: 2010-2021 (Version 0.1.1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.7327711
Coverage in this mirror: south; 2010-07 – 2021-08; 25 km polar stereographic, monthly. Zenodo record(s) total 134 files, 0.28 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/7327711

`obs/satellite/thickness-freeboard-volume/whu_arctic-sit-5km-v1.0/` - A monthly 5 km Arctic sea ice thickness product from 1995 to 2025 using multiple radar altimetry data

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: ERS-2 RA, Envisat RA-2, CryoSat-2 SIRAL
Method: Lead detection, freeboard→SIT, inter-mission bias correction (DTU21 MSS)
Hemisphere / time / resolution: north / 1995-10 – 2025-04 / 5 km polar stereographic, monthly
Version: v1.0
DOI / reference: 10.5281/zenodo.17946273 (version; concept 10.5281/zenodo.12715656)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Xiao, Feng; Zhang, Shengkai; Li, Fei (2025): A monthly 5 km Arctic sea ice thickness product from 1995 to 2025 using multiple radar altimetry data [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17946273
Coverage in this mirror: north; 1995-10 – 2025-04; 5 km polar stereographic, monthly. Zenodo record(s) total 1 files, 1.69 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17946273

`obs/satellite/thickness-freeboard-volume/umd_is2smgpsit-v1/` - IS2SMGPSIT-V1: Daily 25 km Arctic sea ice thickness and volume from fused ICESat-2 and SMOS/SMAP observations, 2018–2025 (v0.1)

Type / level: retrieval product / L4 (daily gridded, fused)
Sensor / platform: ICESat-2 ATLAS + SMOS/SMAP
Method: Gaussian-process fusion of ICESat-2 SIT with SMOS/SMAP thin-ice SIT
Hemisphere / time / resolution: north / 2018-11 – 2025-04 (growth seasons) / 25 km polar stereographic, daily
Version: v1 (record v0.1)
DOI / reference: 10.5281/zenodo.22147361 (version; concept 10.5281/zenodo.22147360)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Petty, Alek (2026): IS2SMGPSIT-V1: Daily 25 km Arctic sea ice thickness and volume from fused ICESat-2 and SMOS/SMAP observations, 2018–2025 (Version v0.1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22147361
Coverage in this mirror: north; 2018-11 – 2025-04 (growth seasons); 25 km polar stereographic, daily. Zenodo record(s) total 54 files, 8.12 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/22147361

`obs/satellite/thickness-freeboard-volume/umd_is2sit-summer-v1/` - Monthly gridded summer Arctic sea ice thickness from ICESat-2, v1

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: ICESat-2 ATLAS (ATL10 R006)
Method: ATL10 freeboard + SnowModel-LG snow (ERA5/MERRA-2), hydrostatic
Hemisphere / time / resolution: north / 2019–2021, May–Aug / 25 km polar stereographic, monthly
Version: v1
DOI / reference: 10.5281/zenodo.18004849 (version; concept 10.5281/zenodo.15375596)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Petty, Alek Aaron (2025): Monthly gridded summer Arctic sea ice thickness from ICESat-2, v1 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.18004849
Coverage in this mirror: north; 2019–2021, May–Aug; 25 km polar stereographic, monthly. Zenodo record(s) total 33 files, 0.27 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/18004849

`obs/satellite/thickness-freeboard-volume/kochi-utokyo_nh-daily-sit-thermo/` - Daily sea ice thickness in the Northern Hemisphere

Type / level: retrieval product / L4 (daily gridded)
Sensor / platform: AMSR-E, AMSR2
Method: SIT from thermodynamic growth history of tracked PMW ice parcels
Hemisphere / time / resolution: north / 2007-01 – 2011-09, 2016-07 – 2025-12 / daily grid (see files)
Version: 2026-06 release
DOI / reference: 10.5281/zenodo.20622229 (version; concept 10.5281/zenodo.17685487)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Kimura, Noriaki; Hasumi, Hiroyasu (2026): Daily sea ice thickness in the Northern Hemisphere [Data set]. Zenodo. https://doi.org/10.5281/zenodo.20622229
Coverage in this mirror: north; 2007-01 – 2011-09, 2016-07 – 2025-12; daily grid (see files). Zenodo record(s) total 31 files, 5.34 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/20622229

`obs/satellite/thickness-freeboard-volume/uhh_smos-smap-sit-v1.0/` - Combined SMOS and SMAP sea ice thickness Arctic (1.0)

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: SMOS MIRAS + SMAP radiometer (L-band)
Method: Combined 40° L-band Tb → thin-ice SIT (Tian-Kunze 2014)
Hemisphere / time / resolution: north / 2015-04 – 2018-04 (Oct–Apr) / 12.5 km EASE2, daily
Version: v1.0
DOI / reference: 10.5281/zenodo.1631856 (version; concept 10.5281/zenodo.1631855)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Schmitt, Amelie; Kaleschke, Lars (2018): Combined SMOS and SMAP sea ice thickness Arctic (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.1631856
Coverage in this mirror: north; 2015-04 – 2018-04 (Oct–Apr); 12.5 km EASE2, daily. Zenodo record(s) total 577 files, 6.20 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/1631856

`obs/satellite/thickness-freeboard-volume/lerma_pmw-sit-ssmi-v1.0/` - PMW Sea Ice Thickness - from SSM/I and SSMIS CDR (v1.0)

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: SSM/I, SSMIS
Method: MLP trained on ICESat-2 SIT using 19/37 GHz Tb
Hemisphere / time / resolution: north / 1992 – 2020 (Oct–Mar) / grid per file
Version: v1.0
DOI / reference: 10.5281/zenodo.13880123 (version; concept 10.5281/zenodo.13880122)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Soriot, Clément; Prigent, Catherine; Jimenez, Carlos (2024): PMW Sea Ice Thickness - from SSM/I and SSMIS CDR (Version v1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.13880123
Coverage in this mirror: north; 1992 – 2020 (Oct–Mar); grid per file. Zenodo record(s) total 1 files, 0.36 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/13880123

`obs/satellite/thickness-freeboard-volume/norce_dasit-v100/` - Drift-aware sea ice thickness maps from satellite remote sensing (v100)

Type / level: retrieval product / L4 (drift-aware)
Sensor / platform: Envisat RA-2, CryoSat-2 + OSI SAF drift/SIC
Method: CCI SIT parcels advected with PMW drift (drift-aware SIT)
Hemisphere / time / resolution: north / Envisat 2002–2012, CS-2 2010–2020 (Oct–Apr) / grid per files
Version: v100 (preliminary)
DOI / reference: 10.5281/zenodo.14733132 (version; concept 10.5281/zenodo.14733131)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Ricker, Robert; Lavergne, Thomas; Hendricks, Stefan; Paul, Stephan; Down, Emily; Killie, Mari Anne; Bocquet, Marion (2025): Drift-aware sea ice thickness maps from satellite remote sensing (Version v100) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.14733132
Coverage in this mirror: north; Envisat 2002–2012, CS-2 2010–2020 (Oct–Apr); grid per files. Zenodo record(s) total 2 files, 16.02 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/14733132

`obs/satellite/thickness-freeboard-volume/ucl_daily-pan-arctic-sit/` - Daily-resolution pan-Arctic sea ice thickness

Type / level: retrieval product / L4 (daily gridded)
Sensor / platform: CryoSat-2, Sentinel-3, ICESat-2
Method: Daily-resolution SIT (Stroeve et al. 2024 method), several algorithms
Hemisphere / time / resolution: north / 2011 – 2022 / daily grid (see files)
Version: 2023-09 release
DOI / reference: 10.5281/zenodo.8366320 (version; concept 10.5281/zenodo.8366319)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Nab, Carmen (2023): Daily-resolution pan-Arctic sea ice thickness [Data set]. Zenodo. https://doi.org/10.5281/zenodo.8366320
Coverage in this mirror: north; 2011 – 2022; daily grid (see files). Zenodo record(s) total 33 files, 8.12 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/8366320

`obs/satellite/thickness-freeboard-volume/tju_ice-3d-ml-sit/` - Arctic Sea Ice Thickness Estimated Using Machine Learning (ICE-3D)

Type / level: retrieval product / L4 (ML estimate)
Sensor / platform: not stated on record
Method: Machine-learning SIT (ICE-3D); record has no description
Hemisphere / time / resolution: north / 2000 – 2023 / not stated
Version: 2026-08 release
DOI / reference: 10.5281/zenodo.22068406 (version; concept 10.5281/zenodo.22068405)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Zheng, Qingyu (2026): Arctic Sea Ice Thickness Estimated Using Machine Learning (ICE-3D) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22068406
Coverage in this mirror: north; 2000 – 2023; not stated. Zenodo record(s) total 24 files, 4.88 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/22068406

`obs/satellite/thickness-freeboard-volume/dmi_cs2-tfmra50-sit-2011-2022-v2/` - Satellite-derived monthly Arctic winter sea ice thickness, snow depth, freeboards, ice draft, and bulk ice density (2011-2022) and validation datasets (2)

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: CryoSat-2 SIRAL (+ ICESat-2)
Method: TFMRA50 radar-freeboard correction; snow depth & SIT retrieval
Hemisphere / time / resolution: north / 2011 – 2022 (winter) / monthly grid (see files)
Version: v2
DOI / reference: 10.5281/zenodo.13145457 (version; concept 10.5281/zenodo.11049390)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Shi, Hoyeon (2024): Satellite-derived monthly Arctic winter sea ice thickness, snow depth, freeboards, ice draft, and bulk ice density (2011-2022) and validation datasets (Version 2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.13145457
Coverage in this mirror: north; 2011 – 2022 (winter); monthly grid (see files). Zenodo record(s) total 1 files, 0.04 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/13145457

`obs/satellite/thickness-freeboard-volume/ucl_s3-narrs-v1.0.0/` - Arctic Snow Depth and Sea Ice Thickness from Sentinel-3 Nadir Radiometer and Radar Synergy (NaRRS) (1.0.0)

Type / level: obs / L2 (along-track)
Sensor / platform: Sentinel-3A/B SRAL + MWR
Method: NaRRS: MWR + SRAL waveform snow depth; SIT hydrostatic
Hemisphere / time / resolution: north (50–81.35°N) / 2016/17 – 2022/23 (Oct–Apr) / along-track, 23.5 km
Version: v1.0.0
DOI / reference: 10.5281/zenodo.17942492 (version; concept 10.5281/zenodo.17942491)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Nelson, Connor (2025): Arctic Snow Depth and Sea Ice Thickness from Sentinel-3 Nadir Radiometer and Radar Synergy (NaRRS) (Version 1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17942492
Coverage in this mirror: north (50–81.35°N); 2016/17 – 2022/23 (Oct–Apr); along-track, 23.5 km. Zenodo record(s) total 2 files, 1.43 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17942492

`obs/satellite/thickness-freeboard-volume/jpl_sits-gnssr-l2-antarctic-v1.0/` - SITS Level-2 Antarctic GNSS-R Sea-Ice Thickness Retrieval Products, Version 1.0 (1.0)

Type / level: obs / L2 (along-track)
Sensor / platform: Spire GNSS-R
Method: GNSS-R reflectivity → along-track SIT
Hemisphere / time / resolution: south / 2022-10 – 2024-01 / along-track
Version: v1.0
DOI / reference: 10.5281/zenodo.23248020 (version; concept 10.5281/zenodo.23248019)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Rodriguez-Alvarez, Nereida; Kim, Seho (2026): SITS Level-2 Antarctic GNSS-R Sea-Ice Thickness Retrieval Products, Version 1.0 (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23248020
Coverage in this mirror: south; 2022-10 – 2024-01; along-track. Zenodo record(s) total 8 files, 0.55 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/23248020

`obs/satellite/thickness-freeboard-volume/jpl_sits-gnssr-l2-arctic-v1.0/` - SITS Level-2 Arctic GNSS-R Sea-Ice Thickness Retrieval Products, Version 1.0 (1.0)

Type / level: obs / L2 (along-track)
Sensor / platform: Spire GNSS-R
Method: GNSS-R reflectivity → along-track SIT
Hemisphere / time / resolution: north / 2022-10 – 2024-01 / along-track
Version: v1.0
DOI / reference: 10.5281/zenodo.21937449 (version; concept 10.5281/zenodo.21937448)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Rodriguez-Alvarez, Nereida; Kim, Seho (2026): SITS Level-2 Arctic GNSS-R Sea-Ice Thickness Retrieval Products, Version 1.0 (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.21937449
Coverage in this mirror: north; 2022-10 – 2024-01; along-track. Zenodo record(s) total 8 files, 0.68 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/21937449

`obs/satellite/thickness-freeboard-volume/norce-awi_dasit-75km-octapr/` - Drift-aware sea ice thickness dataset from satellite remote sensing (trajectories from Oct to Apr and parcel width of 75km)

Type / level: retrieval product / L4 (drift-aware parcels)
Sensor / platform: Envisat RA-2, CryoSat-2 + OSI SAF drift/SIC
Method: DA-SIT customised: Oct–Apr advection, 75 km parcels
Hemisphere / time / resolution: north / Envisat 2002–2012, CS-2 2010–2020 / 75 km parcels (GeoJSON)
Version: v100
DOI / reference: 10.5281/zenodo.20117910 (version; concept 10.5281/zenodo.20117909)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Ricker, Robert; von Albedyll, Luisa (2026): Drift-aware sea ice thickness dataset from satellite remote sensing (trajectories from Oct to Apr and parcel width of 75km) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.20117910
Coverage in this mirror: north; Envisat 2002–2012, CS-2 2010–2020; 75 km parcels (GeoJSON). Zenodo record(s) total 20 files, 14.51 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/20117910

`reanalysis/thickness-freeboard-volume/nersc_topaz4-ml-sit/` - TOPAZ4-ML Sea Ice Thickness (1992-2022) (0.0.1-submitted)

Type / level: reanalysis/model / reanalysis + ML correction
Sensor / platform: TOPAZ4 (assimilates SIC, CS2SMOS) + ML
Method: ML-corrected TOPAZ4 reanalysis thickness
Hemisphere / time / resolution: north / 1992 – 2022 / TOPAZ4 grid (~12.5 km), daily
Version: 0.0.1-submitted
DOI / reference: 10.5281/zenodo.11191854 (version; concept 10.5281/zenodo.11191853)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Edel, Léo; Xie, Jiping; Bertino, Laurent (2024): TOPAZ4-ML Sea Ice Thickness (1992-2022) (Version 0.0.1-submitted) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.11191854
Coverage in this mirror: north; 1992 – 2022; TOPAZ4 grid (~12.5 km), daily. Zenodo record(s) total 32 files, 22.22 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/11191854

Snow on sea ice (冰上积雪)

`reanalysis/snow/nasa_nesosim-v1.1/` - NASA Eulerian Snow On Sea Ice Model Version 1.1 (NESOSIMv1.1) data: 1980 - 2025

Type / level: reanalysis/model / model output (L4)
Sensor / platform: NESOSIM forced by reanalysis snowfall/winds + PMW drift/SIC
Method: Eulerian 2-layer snow budget model
Hemisphere / time / resolution: north / 1980 – 2025 (Sep–Apr) / 100 km, daily
Version: v1.1
DOI / reference: 10.5281/zenodo.18273881 (version; concept 10.5281/zenodo.5164313)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Petty, Alek (2026): NASA Eulerian Snow On Sea Ice Model Version 1.1 (NESOSIMv1.1) data: 1980 - 2025 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.18273881
Coverage in this mirror: north; 1980 – 2025 (Sep–Apr); 100 km, daily. Zenodo record(s) total 45 files, 2.11 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/18273881

`obs/satellite/snow/sysu_snowdepth-pmw-1978-2025/` - A long-term wintertime snow depth dataset on Arctic sea ice (1978-2025) derived from multisource passive microwave radiometer data

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: SMMR, SSM/I, SSMIS
Method: PMW gradient-ratio snow depth with uncertainty
Hemisphere / time / resolution: north / 1978 – 2025 (Oct–Apr) / 25 km polar stereographic, daily
Version: 2026-07 release
DOI / reference: 10.5281/zenodo.21472646 (version; concept 10.5281/zenodo.21472645)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: He, Lian; Peng, Xinning; Zhou, Yi; Hui, Fengming; Chen, Zhuoqi; Chen, Liangbing; Wang, Xianwei; Cheng, Xiao (2026): A long-term wintertime snow depth dataset on Arctic sea ice (1978-2025) derived from multisource passive microwave radiometer data [Data set]. Zenodo. https://doi.org/10.5281/zenodo.21472646
Coverage in this mirror: north; 1978 – 2025 (Oct–Apr); 25 km polar stereographic, daily. Zenodo record(s) total 1 files, 0.88 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/21472646

`reanalysis/snow/eccc_nesosim-mcmc/` - NESOSIM-MCMC Multi-Reanalysis-Average Product With Uncertainty Estimates

Type / level: reanalysis/model / model output (L4)
Sensor / platform: NESOSIM + ERA5/MERRA-2/JRA-55, CloudSat scaling
Method: NESOSIM calibrated by MCMC to snow obs; multi-reanalysis mean + uncertainty
Hemisphere / time / resolution: north / 1980 – 2023 / 100 km, daily
Version: 2025-02 update
DOI / reference: 10.5281/zenodo.14861903 (version; concept 10.5281/zenodo.13307800)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Cabaj, Alex; Petty, Alek A.; Kushner, Paul J. (2025): NESOSIM-MCMC Multi-Reanalysis-Average Product With Uncertainty Estimates [Data set]. Zenodo. https://doi.org/10.5281/zenodo.14861903
Coverage in this mirror: north; 1980 – 2023; 100 km, daily. Zenodo record(s) total 8 files, 4.68 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/14861903

`obs/satellite/snow/sysu_snowdepth-amsr-2002-2024/` - Daily snow depth on Arctic sea ice during the winter season from AMSR-E&2 data (2002 – 2024) (1)

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: AMSR-E, AMSR2
Method: GR(19/7) snow depth with snow-density correction (He et al. 2024)
Hemisphere / time / resolution: north / 2002 – 2024 (Oct–Apr) / 25 km polar stereographic, daily
Version: v1
DOI / reference: 10.5281/zenodo.18251610 (version; concept 10.5281/zenodo.18251609)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: He, Lian; Liu, Lin; Peng, Xinning (2026): Daily snow depth on Arctic sea ice during the winter season from AMSR-E&2 data (2002 – 2024) (Version 1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.18251610
Coverage in this mirror: north; 2002 – 2024 (Oct–Apr); 25 km polar stereographic, daily. Zenodo record(s) total 23 files, 7.65 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/18251610

`obs/satellite/snow/adasa_amsr2-snowdepth-v1/` - Daily gridded Arctic snow depth dataset over sea ice from AMSR2 using an adaptive seasonal algorithm (2012–2023), Version 1

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: AMSR2
Method: Adaptive seasonal algorithm trained on CS-2/IS-2 snow
Hemisphere / time / resolution: north / 2012 – 2023 (freezing seasons) / daily grid (see files)
Version: v1
DOI / reference: 10.5281/zenodo.15695141 (version; concept 10.5281/zenodo.14600500)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Yi Zhou; Xianwei Wang; Ruibo Lei; Yu Zhang; Chentong, Zhang (2025): Daily gridded Arctic snow depth dataset over sea ice from AMSR2 using an adaptive seasonal algorithm (2012–2023), Version 1 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.15695141
Coverage in this mirror: north; 2012 – 2023 (freezing seasons); daily grid (see files). Zenodo record(s) total 1 files, 5.28 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/15695141

`reanalysis/snow/uw_wassail-v1.0.0/` - University of Washington Snow on Antarctic Ice Lagrangian (WASSAIL) model data, v1.0.0 (2003-2025) (1.0.0)

Type / level: reanalysis/model / model output (L4)
Sensor / platform: WASSAIL forced by ERA5 + satellite drift/SIC
Method: Lagrangian parcel snow-budget model
Hemisphere / time / resolution: south / 2003-02 – 2025-02 / 0.75°, daily
Version: v1.0.0
DOI / reference: 10.5281/zenodo.19507962 (version; concept 10.5281/zenodo.19507961)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Campbell, Ethan C. (2026): University of Washington Snow on Antarctic Ice Lagrangian (WASSAIL) model data, v1.0.0 (2003-2025) (Version 1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.19507962
Coverage in this mirror: south; 2003-02 – 2025-02; 0.75°, daily. Zenodo record(s) total 13 files, 16.05 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/19507962

`obs/satellite/snow/legos_laku-snowdepth-v1/` - Multi-frequency altimetry snow depth product over Arctic sea ice (Version 1)

Type / level: retrieval product / L3 (monthly gridded)
Sensor / platform: CryoSat-2 SIRAL (Ku) + ICESat-2 ATLAS
Method: Snow depth = laser minus Ku radar freeboard
Hemisphere / time / resolution: north / 2018 – 2021 (winter) / 25 km, monthly
Version: Version 1
DOI / reference: 10.5281/zenodo.14000765 (version; concept 10.5281/zenodo.14000764)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Carret, Alice; Fleury, Sara (2024): Multi-frequency altimetry snow depth product over Arctic sea ice (Version Version 1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.14000765
Coverage in this mirror: north; 2018 – 2021 (winter); 25 km, monthly. Zenodo record(s) total 21 files, 0.05 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/14000765

`derived-index/snow/awi_snow-clim-w99-amsr2-v1.0/` - Monthly Arctic climatology of snow depth on sea ice from merged passive microwave data and Warren99 Arctic Ocean climatology (1.0)

Type / level: derived index/climatology / L4 climatology
Sensor / platform: AMSR2 + Warren-99 climatology
Method: Merged PMW snow + W99, monthly climatology
Hemisphere / time / resolution: north / climatology (Oct–Apr) / 25 km EPSG:3411, monthly
Version: v1.0
DOI / reference: 10.5281/zenodo.17131889 (version; concept 10.5281/zenodo.17131888)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Hendricks, Stefan (2025): Monthly Arctic climatology of snow depth on sea ice from merged passive microwave data and Warren99 Arctic Ocean climatology (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17131889
Coverage in this mirror: north; climatology (Oct–Apr); 25 km EPSG:3411, monthly. Zenodo record(s) total 7 files, 0.34 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17131889

Drift / deformation (漂移/形变)

`obs/satellite/nsidc_nsidc-0116-v4.1/` - Polar Pathfinder Daily 25 km EASE-Grid Sea Ice Motion Vectors, Version 4 (NSIDC-0116, files v4.1)

Type / level: retrieval product / L3 (daily/weekly gridded)
Sensor / platform: AVHRR, SMMR, SSM/I, SSMIS, AMSR-E + IABP buoys + NCEP winds
Method: Max cross-correlation feature tracking, optimal interpolation with buoys & wind
Hemisphere / time / resolution: both / 1978-11 – latest year / 25 km EASE-Grid, daily/weekly
Version: v4.1
DOI / reference: 10.5067/INAWUWO7QH7B
Licence: NASA open data policy, no restrictions; please cite.
Citation: Tschudi, M., Meier, W. N., Stewart, J. S., Fowler, C., Maslanik, J. (2019). Polar Pathfinder Daily 25 km EASE-Grid Sea Ice Motion Vectors, Version 4. NASA NSIDC DAAC. doi:10.5067/INAWUWO7QH7B.
Coverage in this mirror: NH + SH, 25 km EASE-Grid, 1978-11 to latest available year: yearly NetCDF files of daily vectors (daily/) and of weekly means (weekly/).
Subsetting decision: Full record taken (yearly NetCDF, ~3 GB total). browse/ images and input_sources/ (per-sensor input vectors) were skipped. No provider checksums on the server; MD5 computed after opening each NetCDF.
Source: https://daacdata.apps.nsidc.org/pub/DATASETS/nsidc0116_icemotion_vectors_v4/ (Earthdata login)

`obs/satellite/drift-deformation/mcgill-awi_rgps-deformation-composite-v1.0/` - Deformation composite of the RADARSAT Geophysical Processor System (RGPS) Lagrangian motion data (1.0)

Type / level: retrieval product / L3 (3-day gridded)
Sensor / platform: RADARSAT-1 SAR (RGPS)
Method: Strain rates from RGPS Lagrangian trajectories, composited
Hemisphere / time / resolution: north / 1997 – 2008 (Jan–Mar) / 10 km, 3-day
Version: v1.0
DOI / reference: 10.5281/zenodo.6321327 (version; concept 10.5281/zenodo.6321326)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Bouchat, Amélie; Hutter, Nils (2022): Deformation composite of the RADARSAT Geophysical Processor System (RGPS) Lagrangian motion data (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.6321327
Coverage in this mirror: north; 1997 – 2008 (Jan–Mar); 10 km, 3-day. Zenodo record(s) total 12 files, 2.96 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/6321327

`obs/satellite/drift-deformation/eccc_sidrr-asits-v1.3/` - Sea ice deformation and rotation rates (SIDRR) from the ECCC-ASITS

Type / level: retrieval product / L2/L3 (daily, SAR triangles)
Sensor / platform: Sentinel-1 SAR, RADARSAT Constellation
Method: ECCC-ASITS SAR motion vectors → triangulated deformation/rotation
Hemisphere / time / resolution: north / 2017 – 2025-08 / SAR triangles, daily
Version: v1.3
DOI / reference: 10.5281/zenodo.18088256 (version; concept 10.5281/zenodo.11520802)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Plante, Mathieu; Lemieux, Jean-François; Tremblay, L. Bruno; Bouchat, Amélie; Ringeisen, Damien; Blain, Philippe; Howell, Stephen; Brady, Mike; Alexander, Komarov; Duval, Beatrice; Yakuden, Lekima; Labelle, Frédérique (2025): Sea ice deformation and rotation rates (SIDRR) from the ECCC-ASITS [Data set]. Zenodo. https://doi.org/10.5281/zenodo.18088256
Coverage in this mirror: north; 2017 – 2025-08; SAR triangles, daily. Zenodo record(s) total 9 files, 13.14 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/18088256

`obs/satellite/drift-deformation/nsidc_sar-antarctic-velocity-v2/` - SAR-derived Antarctic sea ice velocity (2015-2023) (v2)

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: Sentinel-1 SAR
Method: Feature tracking (nansencenter sea_ice_drift) on S1 pairs
Hemisphere / time / resolution: south / 2015 – 2023 / EASE-Grid South, daily
Version: v2
DOI / reference: 10.5281/zenodo.22942305 (version; concept 10.5281/zenodo.18737892)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Koo, YoungHyun (2026): SAR-derived Antarctic sea ice velocity (2015-2023) (Version v2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22942305
Coverage in this mirror: south; 2015 – 2023; EASE-Grid South, daily. Zenodo record(s) total 9 files, 0.24 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/22942305

`obs/satellite/drift-deformation/utokyo_nh-daily-ice-velocity/` - Gridded Daily Sea Ice Velocity in the Northern Hemisphere

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: AMSR-E, AMSR2 (36/18 GHz)
Method: Maximum cross-correlation, 60 km template
Hemisphere / time / resolution: north / 2007 – 2025 / 60 km, daily
Version: 2025-11 release
DOI / reference: 10.5281/zenodo.17694536 (version; concept 10.5281/zenodo.17694535)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Kimura, Noriaki (2025): Gridded Daily Sea Ice Velocity in the Northern Hemisphere [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17694536
Coverage in this mirror: north; 2007 – 2025; 60 km, daily. Zenodo record(s) total 6 files, 0.21 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17694536

`obs/satellite/drift-deformation/sysu_amsr2-summer-velocity-v1.0/` - Arctic sea ice velocity in summer from AMSR2 (2013-2023) (v1.0)

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: AMSR2 36.5 GHz
Method: Continuous max cross-correlation on polarisation difference
Hemisphere / time / resolution: north / 2013 – 2023 (May–Sep) / daily grid (see files)
Version: v1.0
DOI / reference: 10.5281/zenodo.12618229 (version; concept 10.5281/zenodo.12618228)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Shi, Qian; Yang, Qinghua (2024): Arctic sea ice velocity in summer from AMSR2 (2013-2023) (Version v1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.12618229
Coverage in this mirror: north; 2013 – 2023 (May–Sep); daily grid (see files). Zenodo record(s) total 12 files, 0.85 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/12618229

`derived-index/drift-deformation/gsfc_lagrangian-parcel-tracks-v1.0/` - Daily Lagrangian Arctic sea ice parcel drift tracks with coincident ice and atmospheric conditions (Part 1) (1.0)

Type / level: derived index/climatology / L4 (Lagrangian database)
Sensor / platform: PMW drift/SIC + reanalysis
Method: Daily parcel drift tracks with coincident ice/atmosphere conditions
Hemisphere / time / resolution: north / 2002 – 2020 (sea-ice years) / daily trajectories
Version: v1.0 (parts 1, 2, 2a, 3)
DOI / reference: 10.5281/zenodo.6646932 (version; concept 10.5281/zenodo.6646931); 10.5281/zenodo.6656294 (version; concept 10.5281/zenodo.6656293); 10.5281/zenodo.6647508 (version; concept 10.5281/zenodo.6647507); 10.5281/zenodo.6647807 (version; concept 10.5281/zenodo.6647806)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Horvath, Sean; Boisvert, Linette; Parker, Chelsea; Webster, Melinda; Taylor, Patrick; Boeke, Robin (2022): Daily Lagrangian Arctic sea ice parcel drift tracks with coincident ice and atmospheric conditions (Part 1) (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.6646932 | Horvath, Sean; Boisvert, Linette; Parker, Chelsea; Webster, Melinda; Taylor, Patrick; Boeke, Robin (2022): Daily Lagrangian Arctic sea ice parcel drift tracks with coincident ice and atmospheric conditions (Part 2a) (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.6656294 | Horvath, Sean; Boisvert, Linette; Parker, Chelsea; Webster, Melinda; Taylor, Patrick; Boeke, Robin (2022): Daily Lagrangian Arctic sea ice parcel drift tracks with coincident ice and atmospheric conditions (Part 2) (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.6647508 | Horvath, Sean; Boisvert, Linette; Parker, Chelsea; Webster, Melinda; Taylor, Patrick; Boeke, Robin (2022): Daily Lagrangian Arctic sea ice parcel drift tracks with coincident ice and atmospheric conditions (Part 3) (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.6647807
Coverage in this mirror: north; 2002 – 2020 (sea-ice years); daily trajectories. Zenodo record(s) total 74 files, 130.29 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/6646932 https://zenodo.org/records/6656294 https://zenodo.org/records/6647508 https://zenodo.org/records/6647807

Age / type (冰龄/冰型)

`obs/satellite/age-type/nersc_sea-ice-age-cdr-v2.1/` - Arctic Sea Ice Age Climate Data Record Version 2.1 (monthly) (2.1.2)

Type / level: retrieval product / L4 (monthly gridded)
Sensor / platform: PMW via OSI SAF SIC + drift CDRs
Method: Lagrangian advection of ice-age fractions with OSI SAF drift/SIC
Hemisphere / time / resolution: north / 1991 – 2024 / monthly grid (see files)
Version: v2.1.2
DOI / reference: 10.5281/zenodo.17691212 (version; concept 10.5281/zenodo.15773500)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Korosov, Anton (2025): Arctic Sea Ice Age Climate Data Record Version 2.1 (monthly) (Version 2.1.2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17691212
Coverage in this mirror: north; 1991 – 2024; monthly grid (see files). Zenodo record(s) total 34 files, 0.09 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17691212

`obs/satellite/age-type/nsidc_ease-iceage-early-v4.1/` - Early EASE-GRID Sea Ice Age, 1978-1983 (4.1_early)

Type / level: retrieval product / L4 (weekly gridded)
Sensor / platform: PMW + buoys (NSIDC-0116 motion)
Method: Lagrangian tracking, annual age classes (spin-up period)
Hemisphere / time / resolution: north / 1978-10 – 1983 / 12.5 km EASE-Grid, weekly
Version: v4.1_early
DOI / reference: 10.5281/zenodo.7659077 (version; concept 10.5281/zenodo.7659076)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Meier, Walter N.; Stewart, J. Scott (2023): Early EASE-GRID Sea Ice Age, 1978-1983 (Version 4.1_early) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.7659077
Coverage in this mirror: north; 1978-10 – 1983; 12.5 km EASE-Grid, weekly. Zenodo record(s) total 6 files, 0.02 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/7659077

`obs/satellite/age-type/utokyo_nh-daily-sea-ice-age/` - Daily sea ice age data for the entire Northern Hemisphere

Type / level: retrieval product / L4 (daily)
Sensor / platform: AMSR-E, AMSR2
Method: Backward tracking with AMSR velocity + SIC → mean/oldest age
Hemisphere / time / resolution: north / 2007 – 2022 / daily, AMSR grid (text)
Version: 2025-11 release
DOI / reference: 10.5281/zenodo.17743866 (version; concept 10.5281/zenodo.17743865)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Kimura, Noriaki; Hasumi, Hiroyasu (2025): Daily sea ice age data for the entire Northern Hemisphere [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17743866
Coverage in this mirror: north; 2007 – 2022; daily, AMSR grid (text). Zenodo record(s) total 12 files, 4.35 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17743866

`obs/satellite/age-type/sysu_arctic-ice-type-2002-2021/` - Arctic Sea Ice Type Products Based on Active and Passive Microwave Remote Sensing (2002-2021)

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: FY-3 MWRI + other PMW + scatterometer
Method: K-means clustering → open water / FYI / MYI
Hemisphere / time / resolution: north / 2002 – 2021 / 12.5 km, daily
Version: 2022-10 release
DOI / reference: 10.5281/zenodo.10889942 (version; concept 10.5281/zenodo.10889941)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: 李, 锦鑫; 叶, 玉芳; 胡, 海涵 (2022): Arctic Sea Ice Type Products Based on Active and Passive Microwave Remote Sensing (2002-2021) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10889942
Coverage in this mirror: north; 2002 – 2021; 12.5 km, daily. Zenodo record(s) total 1 files, 0.06 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/10889942

Melt / albedo / temperature / surface (融化/反照率/温度/表面)

`obs/satellite/melt-albedo-temperature/gsfc_ahra-melt-onset-v5/` - Arctic sea ice snow melt onset dates from the Advanced Horizontal Range Algorithm, version 5 (1979 - 2022)

Type / level: retrieval product / L3 (annual gridded)
Sensor / platform: SMMR, SSM/I, SSMIS
Method: Advanced Horizontal Range Algorithm on PMW Tb time series
Hemisphere / time / resolution: north / 1979 – 2022 / 25 km, annual
Version: v5 (v05r00)
DOI / reference: 10.5281/zenodo.10137555 (version; concept 10.5281/zenodo.10137554)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Bliss, Angela C. (2023): Arctic sea ice snow melt onset dates from the Advanced Horizontal Range Algorithm, version 5 (1979 - 2022) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10137555
Coverage in this mirror: north; 1979 – 2022; 25 km, annual. Zenodo record(s) total 52 files, 0.02 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/10137555

`obs/satellite/melt-albedo-temperature/gapfree-arctic-albedo-v3/` - Gap-Free Arctic Sea Ice Albedo Dataset (2000–2025; Daily, May–September), Version 3

Type / level: retrieval product / L4 (daily gap-free)
Sensor / platform: MODIS Terra + PMW Tb
Method: ML albedo from MODIS (melt-pond aware); gaps from albedo–Tb model
Hemisphere / time / resolution: north (>60°N) / 2000 – 2024 (May–Sep) / 12.5 km polar stereographic, daily
Version: v3
DOI / reference: 10.5281/zenodo.23202177 (version; concept 10.5281/zenodo.15307662)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Zhao, Ziwei; Ding, Yinghui; Wang, Mengsi; Li, Xijia; Qu, Ying (2026): Gap-Free Arctic Sea Ice Albedo Dataset (2000–2025; Daily, May–September), Version 3 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23202177
Coverage in this mirror: north (>60°N); 2000 – 2024 (May–Sep); 12.5 km polar stereographic, daily. Zenodo record(s) total 25 files, 1.13 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/23202177

`obs/satellite/melt-albedo-temperature/dmi_t2m-arctic-seaice-v0.1/` - Monthly near-surface air temperatures over Arctic sea ice based on satellite observations

Type / level: retrieval product / L4 (monthly gap-free)
Sensor / platform: AVHRR IST CDR (DMI)
Method: T2m from sea/ice surface temperature CDR via empirical model
Hemisphere / time / resolution: north / 1982-01 – 2021-05 / 0.05°, monthly
Version: v0.1
DOI / reference: 10.5281/zenodo.10555223 (version; concept 10.5281/zenodo.10555222)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Nielsen-Englyst, Pia; Høyer, Jacob (2024): Monthly near-surface air temperatures over Arctic sea ice based on satellite observations [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10555223
Coverage in this mirror: north; 1982-01 – 2021-05; 0.05°, monthly. Zenodo record(s) total 1 files, 3.62 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/10555223

`obs/satellite/melt-albedo-temperature/ouc_arctic-freeze-onset/` - Arctic Sea Ice Freeze Onsest (1)

Type / level: retrieval product / L3 (annual gridded)
Sensor / platform: AMSR-E, AMSR2 + NT2 SIC
Method: Improved PMW freeze-onset algorithm
Hemisphere / time / resolution: north / 2003 – 2021 (no 2011–12) / 12.5 km, annual
Version: v1
DOI / reference: 10.5281/zenodo.8014215 (version; concept 10.5281/zenodo.8014214)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Qu, zhifeng; Su, Jie (2023): Arctic Sea Ice Freeze Onsest (Version 1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.8014215
Coverage in this mirror: north; 2003 – 2021 (no 2011–12); 12.5 km, annual. Zenodo record(s) total 17 files, 0.15 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/8014215

`obs/satellite/melt-albedo-temperature/sysu_ascat-form-drag/` - Wintertime daily Arctic sea ice surface form drag from ASCAT during 2006-2021

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: MetOp ASCAT
Method: Surface form drag from scatterometer backscatter (SIC>90%)
Hemisphere / time / resolution: north / 2006 – 2021 (Nov–Apr) / 12.5 km, daily
Version: 2023-12 release
DOI / reference: 10.5281/zenodo.10421427 (version; concept 10.5281/zenodo.10421426)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Zhang, Zhilun (2023): Wintertime daily Arctic sea ice surface form drag from ASCAT during 2006-2021 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10421427
Coverage in this mirror: north; 2006 – 2021 (Nov–Apr); 12.5 km, daily. Zenodo record(s) total 1 files, 1.84 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/10421427

`reanalysis/melt-albedo-temperature/ncar_era5-skt-biascorr/` - Bias corrected era5 skin temperature over the Arctic sea ice – 1981 to 2018 monthly means and climatology

Type / level: reanalysis/model / reanalysis, bias-corrected
Sensor / platform: ERA5 + ML correction
Method: ML correction of winter clear-sky skin-temperature bias
Hemisphere / time / resolution: north / 1981 – 2018 / ERA5 grid, monthly + climatology
Version: 2023-05 release
DOI / reference: 10.5281/zenodo.8338265 (version; concept 10.5281/zenodo.8338264)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Zampieri, Lorenzo (2023): Bias corrected era5 skin temperature over the Arctic sea ice – 1981 to 2018 monthly means and climatology [Data set]. Zenodo. https://doi.org/10.5281/zenodo.8338265
Coverage in this mirror: north; 1981 – 2018; ERA5 grid, monthly + climatology. Zenodo record(s) total 39 files, 1.22 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/8338265

`obs/satellite/melt-albedo-temperature/sysu_quikscat-form-drag/` - Wintertime daily Arctic sea ice surface form drag from QuikSCAT during 1999-2009

Type / level: retrieval product / L3 (daily gridded)
Sensor / platform: QuikSCAT SeaWinds
Method: Surface form drag from scatterometer backscatter (SIC>90%)
Hemisphere / time / resolution: north / 1999 – 2009 (Nov–Apr) / 12.5 km, daily
Version: 2023-12 release
DOI / reference: 10.5281/zenodo.10421183 (version; concept 10.5281/zenodo.10421182)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Zhang, Zhilun (2023): Wintertime daily Arctic sea ice surface form drag from QuikSCAT during 1999-2009 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.10421183
Coverage in this mirror: north; 1999 – 2009 (Nov–Apr); 12.5 km, daily. Zenodo record(s) total 1 files, 1.12 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/10421183

`obs/satellite/melt-albedo-temperature/whu_antarctic-albedo-viirs-1km-uncertainty-v1/` - Uncertainty of the daily 1 km seamless Antarctic sea ice albedo product (v1)

Type / level: retrieval product / L4 (daily uncertainty)
Sensor / platform: S-NPP VIIRS
Method: Monte Carlo + error propagation uncertainty of the 1 km albedo
Hemisphere / time / resolution: south / 2012 – 2021 / 1 km, daily
Version: v1
DOI / reference: 10.5281/zenodo.15067607 (version; concept 10.5281/zenodo.15067606)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Ma, Chao; Hao, Weifeng; Cheng, Qing; Ye, Fan; Wu, Haojian; Li, Fei (2025): Uncertainty of the daily 1 km seamless Antarctic sea ice albedo product (Version v1) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.15067607
Coverage in this mirror: south; 2012 – 2021; 1 km, daily. Zenodo record(s) total 10 files, 4.66 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/15067607

`obs/satellite/melt-albedo-temperature/whu_antarctic-albedo-viirs-1km-v2/` - Daily 1 km seamless Antarctic sea ice albedo product from VIIRS data (v2)

Type / level: retrieval product / L4 (daily seamless)
Sensor / platform: S-NPP VIIRS
Method: MBRI albedo retrieval; cloudy pixels reconstructed
Hemisphere / time / resolution: south / 2012 – 2021 / 1 km, daily
Version: v2
DOI / reference: 10.5281/zenodo.11216156 (version; concept 10.5281/zenodo.11196870)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Ma, Chao; Hao, Weifeng; Cheng, Qing; Ye, Fan; Qu, Ying; Gao, Shengjun; Wu, Haojian; Li, Fei (2024): Daily 1 km seamless Antarctic sea ice albedo product from VIIRS data (Version v2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.11216156
Coverage in this mirror: south; 2012 – 2021; 1 km, daily. Zenodo record(s) total 10 files, 49.83 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/11216156

Leads / polynya / landfast / iceberg (冰间水道/冰间湖/固定冰/冰山)

`obs/satellite/leads-polynya-landfast-iceberg/sysu_deep-aa-polynya-v1.1/` - Daily Edge of Each Polynya in Antarctic (DEEP-AA) (1.1.0-2023)

Type / level: retrieval product / L4 (daily polynya maps)
Sensor / platform: AMSR-E/AMSR2 SIC & PSSM
Method: Polynya identification/tracking on SIC thresholds (use SIC60_6.25km_20d)
Hemisphere / time / resolution: south / 2003 – 2022 (Apr–Oct) / 6.25 km, daily
Version: v1.1.0-2023
DOI / reference: 10.5281/zenodo.13358042 (version; concept 10.5281/zenodo.11379147)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Lin, Yichen; Nakayama, Yoshihiro; Liang, Kaixin; Huang, Yongtao; Chen, Dake; Yang, Qinghua (2024): Daily Edge of Each Polynya in Antarctic (DEEP-AA) (Version 1.1.0-2023) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.13358042
Coverage in this mirror: south; 2003 – 2022 (Apr–Oct); 6.25 km, daily. Zenodo record(s) total 29 files, 5.20 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/13358042

`obs/satellite/leads-polynya-landfast-iceberg/fio-ouc_arctic-landfast-weekly/` - 25-km gridded weekly Arctic landfast sea ice distribution from 1976-2026

Type / level: retrieval product / L3 (weekly gridded, ice-chart analysis)
Sensor / platform: US NIC ice charts (multi-sensor analyst)
Method: Landfast ice extracted from NIC weekly/biweekly charts
Hemisphere / time / resolution: north / 1976 – 2026 / 25 km, weekly
Version: 2026-04 release
DOI / reference: 10.5281/zenodo.19546997 (version; concept 10.5281/zenodo.19546996)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Jiechen, Zhao (2026): 25-km gridded weekly Arctic landfast sea ice distribution from 1976-2026 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.19546997
Coverage in this mirror: north; 1976 – 2026; 25 km, weekly. Zenodo record(s) total 4 files, 0.02 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/19546997

`obs/satellite/leads-polynya-landfast-iceberg/uwm_viirs-leads-unet/` - VIIRS Sea ice leads detections using a U-Net

Type / level: retrieval product / L3 (daily lead maps)
Sensor / platform: S-NPP VIIRS thermal IR
Method: U-Net CNN lead detection
Hemisphere / time / resolution: north / 2011 – 2021 (Nov–Apr) / ~1 km, daily
Version: 2022-09 release
DOI / reference: 10.5281/zenodo.7046261 (version; concept 10.5281/zenodo.7046260)
Licence: CC0 1.0 Universal (Zenodo record licence); redistribution allowed.
Citation: Hoffman, Jay; Ackerman, Steven; Liu, Yinghui; Key, Jeffrey; McConnell, Iain (2022): VIIRS Sea ice leads detections using a U-Net [Data set]. Zenodo. https://doi.org/10.5061/dryad.1vhhmgqwd
Coverage in this mirror: north; 2011 – 2021 (Nov–Apr); ~1 km, daily. Zenodo record(s) total 13 files, 16.11 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/7046261

`obs/satellite/leads-polynya-landfast-iceberg/sxu-sysu_antarctic-icebergs-s1/` - A Six-year circum-Antarctic icebergs dataset (2018-2023)

Type / level: retrieval product / L3 (annual iceberg vectors)
Sensor / platform: Sentinel-1 SAR
Method: Incremental random-forest iceberg detection + manual correction
Hemisphere / time / resolution: south (<55°S) / 2018-10 – 2023-10 / vectors ≥0.04 km², annual
Version: 2025-09 release
DOI / reference: 10.5281/zenodo.17165466 (version; concept 10.5281/zenodo.14588507)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Liu, xuying; Chen, zilong (2025): A Six-year circum-Antarctic icebergs dataset (2018-2023) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17165466
Coverage in this mirror: south (<55°S); 2018-10 – 2023-10; vectors ≥0.04 km², annual. Zenodo record(s) total 3 files, 0.69 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/17165466

`obs/satellite/leads-polynya-landfast-iceberg/gsfc_is2-lead-fraction-v1.0/` - ICESat-2 lead fraction and chord length estimates (1.0)

Type / level: retrieval product / L2/L3 (along-track + gridded)
Sensor / platform: ICESat-2 ATLAS (ATL07)
Method: Lead fraction & chord length from ATL07 surface type
Hemisphere / time / resolution: both / NH 2018-11–2019-04, SH 2019-05–2019-09 / along-track + 25 km
Version: v1.0
DOI / reference: 10.5281/zenodo.4553552 (version; concept 10.5281/zenodo.4075390)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Petty, Alek (2021): ICESat-2 lead fraction and chord length estimates (Version 1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.4553552
Coverage in this mirror: both; NH 2018-11–2019-04, SH 2019-05–2019-09; along-track + 25 km. Zenodo record(s) total 14 files, 0.46 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/4553552

`obs/satellite/leads-polynya-landfast-iceberg/uwm_modis-leads-unet/` - MODIS Sea ice leads detections using a U-Net

Type / level: retrieval product / L3 (daily lead maps)
Sensor / platform: Terra/Aqua MODIS thermal IR
Method: U-Net CNN lead detection
Hemisphere / time / resolution: north / 2002 – 2021 (Nov–Apr) / ~1 km, daily
Version: 2022-07 release
DOI / reference: 10.5281/zenodo.6904022 (version; concept 10.5281/zenodo.6904021)
Licence: CC0 1.0 Universal (Zenodo record licence); redistribution allowed.
Citation: Hoffman, Jay; Ackerman, Steven; Liu, Yinghui; Key, Jeffrey; McConnell, Iain (2022): MODIS Sea ice leads detections using a U-Net [Data set]. Zenodo. https://doi.org/10.5061/dryad.79cnp5hz2
Coverage in this mirror: north; 2002 – 2021 (Nov–Apr); ~1 km, daily. Zenodo record(s) total 22 files, 35.35 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/6904022

Brightness temperature (亮温)

`obs/satellite/brightness-temperature/uhh_smap-l1b-tb-arctic/` - SMAP L1B Brightness Temperatures Arctic (3.0)

Type / level: obs / L1B → gridded daily Tb
Sensor / platform: SMAP radiometer (L-band)
Method: SMAP L1B Tb gridded (40° incidence)
Hemisphere / time / resolution: north / 2015-04 – 2018-05 / 12.5 km EASE2, daily
Version: 2018-11 release
DOI / reference: 10.5281/zenodo.1627380 (version; concept 10.5281/zenodo.1627379)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Schmitt, Amelie; Kaleschke, Lars (2018): SMAP L1B Brightness Temperatures Arctic (Version 3.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.1627380
Coverage in this mirror: north; 2015-04 – 2018-05; 12.5 km EASE2, daily. Zenodo record(s) total 1150 files, 11.27 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/1627380

`obs/satellite/brightness-temperature/uhh_smos-tb40-arctic-v620/` - SMOS Brightness Temperatures at 40° incidence angle Arctic (v620)

Type / level: obs / L1C → gridded daily Tb
Sensor / platform: SMOS MIRAS (L-band)
Method: Multi-angle fit to 40° incidence, gridded
Hemisphere / time / resolution: north / 2015-04 – 2018-05 / 12.5 km EASE2, daily
Version: v620
DOI / reference: 10.5281/zenodo.1631762 (version; concept 10.5281/zenodo.1631761)
Licence: Creative Commons Attribution 4.0 International (Zenodo record licence); redistribution allowed with attribution.
Citation: Schmitt, Amelie; Kaleschke, Lars (2018): SMOS Brightness Temperatures at 40° incidence angle Arctic (Version v620) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.1631762
Coverage in this mirror: north; 2015-04 – 2018-05; 12.5 km EASE2, daily. Zenodo record(s) total 1154 files, 13.57 GB.
Subsetting decision: Full Zenodo record(s) mirrored, files MD5-checked against the Zenodo API. ZIP/TAR archives whose contents are NetCDF/GeoTIFF/CSV are unpacked (members stored unchanged under a folder named after the archive); other archives (7z, rar, multi-part zips, nested or mixed contents) are kept as-is.
Source: https://zenodo.org/records/1631762

whale12 — Tue, 13 Oct 2026 04:34:08 GMT

已完成实验归档

更新时间：2026-10-19T10:53:34.163877+08:00。本数据集仅收录训练与正式评测均已完成的实验；基础模型只需完成正式评测。当前 20组实验、4476条逐题轨迹，不重复计算309题的固定96子集。

实验索引

实验目录（模型 / 算法 / 版本）	训练完成	正式评测完成
Qwen3.8-27B/GRPO/step100	100/100	fixed96 · 96/96
Qwen3.8-27B/REINFORCE_plus_plus/step100	100/100	fixed96 · 96/96
Qwen3.8-27B/ReMax/step100	100/100	fixed96 · 96/96
Qwen3.8-27B/RLOO/step100	100/100	fixed96 · 96/96
Muse-Glimmer-30B/MASC/historical-scorefix01-step064	64/64	fixed96 · 96/96
Muse-Glimmer-30B/GRPO/step100	100/100	fixed96 · 96/96
Muse-Glimmer-30B/SFT/clear199-tasks100-1epoch-step725	725/725（1epoch）	full309 · 309/309（含固定96子集）
Qwen3.8-27B/SFT/clear199-tasks100-1epoch-step725	725/725（1epoch）	full309 · 309/309（含固定96子集）
Muse-Glimmer-30B/MASC-v2/step100	100/100	full309 · 309/309（含固定96子集）
Muse-Glimmer-30B/MASC/corrector-credit-step100	100/100	full309 · 309/309（含固定96子集）
Qwen3.8-27B/MASC/corrector-credit-step100	100/100	full309 · 309/309（含固定96子集）
Qwen3.8-27B/MASC-v2/step100	100/100	full309 · 309/309（含固定96子集）
Muse-Glimmer-30B/ReMax/step100	100/100	full309 · 309/309（含固定96子集）
Muse-Glimmer-30B/REINFORCE_plus_plus/step100	100/100	full309 · 309/309（含固定96子集）
Muse-Glimmer-30B/RLOO/step100	100/100	full309 · 309/309（含固定96子集）
Muse-Glimmer-30B/MASC/no-corrector-credit-step100	100/100	fixed96 · 96/96
Qwen3.8-27B/MASC/no-corrector-credit-step100	100/100	fixed96 · 96/96
Muse-Glimmer-30B/Base/base-model	基础模型，无训练	full309 · 309/309
Muse-Glimmer-30B/Summary/base-model	基础模型，无训练	full309 · 309/309
Muse-Glimmer-30B/SAGE/base-model	基础模型，无训练	full309 · 309/309

统一结构

模型/算法/实验版本/
├── README.md
├── completion.json
├── manifest.sha256.json
├── training/                 # 训练代码、配置、脚本、日志
├── checkpoints/step100/      # 或step64；仅已完成训练
└── evaluation/
    ├── fixed96/             # 正式96题，或从full309提取的子集统计
    └── full309/             # 正式309题；仅在实际完成时存在
        ├── tasks/           # 场景/task编号：结果、交互、工具调用和视觉记录
        ├── logs/
        ├── config/
        └── trajectory_index.json

基础模型实验没有 training/ 或 checkpoints/；共享Muse基座保存在 shared_models/Muse-Glimmer-30B。Qwen训练包为LoRA及续训状态，需配合训练配置指定的Qwen基座，不包含另一个独立完整Qwen基础权重副本。

结果和轨迹在哪里

逐题JSON包含 dialogue 等交互字段，并非只有分数；Base/Summary还包含 tool_calls。vision.json、timing.log、服务日志和JSONL计时事件按原始保存情况归档。每组的 trajectory_index.json 可逐题定位，manifest提供SHA256。没有记录的原始HTTP报文无法事后补造。

完成标准：训练达到该实验原定目标，正式评测任务集合完整。任务判定失败或协议失败会保留，不因分数低而删除。历史Muse MASC目标为64步；Qwen旧版MASC为100步，不与仍在运行的新版纠错MASC混用。

不在本次归档范围

仍在训练或正式评测尚未完整结束的实验，不上传临时权重或部分测试结果。Qwen基础SAGE的309题仅有文档汇总，当前未找到逐题原始文件，未列入完整归档。

目录已统一，旧目录对应关系见 migration_map.json。provenance中的旧清单仅作历史记录，不作为新版文件路径。迁移保留已有权重、逐题结果、轨迹和日志的内容哈希，新增已完成Qwen旧版MASC100权重及复现材料。传输不使用HTTP代理。GitHub不存放本数据集的原始产物。

Scallions — Mon, 12 Oct 2026 15:27:33 GMT

Greenland GPS PWV — pilot year 2020

数据进度与范围 / Status & Coverage

最后更新 (last updated): 2026-10-21 05:47 CST (UTC+8) 试点年份 (pilot year) 2020；区域 (region) 55–84°N, 80–10°W

部分 (part)	状态 (status)	时间范围 (temporal)	空间/分辨率 (spatial)	内容/变量 (variables)	文件数 / 大小 (files / size)
NGL GNSS ZTD (`raw/ngl/`)	✅ 完成 (done)	2020 全年 (yearly zips)	205 个候选测站, 124 个有 2020 年数据 (stations with 2020 data)	5 分钟 ZTD、梯度 (5-min tropo)	121 / 0.22 GB
ERA5 单层 (`raw/era5/era5_sl_*`)	✅ 完成 (done) 12/12 (100.0%) 月	计划 2020-01 – 2020-12；已上传 2020-01 – 2020-12	区域子集 (regional subset)	地面气压/温度/TCWV 等 (single-level fields for ZHD & comparison)	12 / 1.29 GB
ERA5 气压层 (`raw/era5/era5_pl_*`)	⏳ 进行中 (in progress) 5/12 (41.7%) 月	计划 2020-01 – 2020-12；已上传 2020-01 – 2020-05	0.5°, 3 小时 (3-hourly), 27 层 1000–100 hPa	t, q, z（用于 Tm）	5 / 1.26 GB
IGRA2 探空 (`raw/igra2/`)	✅ 完成 (done)	全记录 (period of record)	12 个探空站 (radiosonde stations)	data-por + derived-por	24 / 1.08 GB
逐小时 PWV (`processed/pwv_hourly/`)	⏳ 进行中 (in progress) 5/12 (41.7%) 月已处理 (months processed)	已处理月份 (processed months): 2020-01, 2020-02, 2020-03, 2020-04, 2020-05	119 个测站 (stations)	PWV (mm) + 质量标记 (QC flags)	本地已算，全年处理完后统一上传 (computed locally, uploaded after finalize)

与 ERA5 TCWV、IGRA2 的对比结果在 12 个月处理完后生成于 processed/comparison/（目前为基于 NGL 自带 PWV 的临时对比 provisional）。

Hourly GNSS precipitable water vapour (PWV) for ~119 GNSS stations in Greenland, Iceland and the Canadian Arctic (55–84°N, 80–10°W), derived from Nevada Geodetic Laboratory 5-min zenith total delays, with ERA5-based ZHD (Saastamoinen) and Tm, and comparisons against ERA5 TCWV and IGRA2 radiosonde PW. See METHODS.md.

Layout

raw/ngl/ NGL IGS20 tropo yearly zips (+ station list files), raw/era5/ monthly ERA5 NetCDF (single + pressure levels), raw/igra2/ IGRA2 data-por / derived-por zips. Full manifest with md5 and source URLs: manifest_raw.csv.
processed/pwv_hourly/ per-station hourly PWV with quality flags; processed/igra2/ 2020 radiosonde PW; processed/comparison/ stats tables and figures.
stations.csv station list (name, lat, lon, ellipsoidal height, data span).

Sources, licences, citations

GNSS ZTD: Nevada Geodetic Laboratory, http://geodesy.unr.edu. Cite: Blewitt, G., Hammond, W. C., Kreemer, C. (2018), Harnessing the GPS data explosion for interdisciplinary science, Eos 99, doi:10.1029/2018EO104623. Freely available; please acknowledge NGL and the original GNSS data providers (e.g. GNET/DTU Space & SDFE, IMO Iceland, NRCan).
ERA5: Hersbach, H. et al. (2020), The ERA5 global reanalysis, QJRMS 146, 1999–2049, doi:10.1002/qj.3803. Data: Copernicus Climate Change Service (C3S) Climate Data Store, doi:10.24381/cds.adbb2d47 and doi:10.24381/cds.bd0915c6. Licence CC-BY 4.0. "Contains modified Copernicus Climate Change Service information 2020."
IGRA2: Durre, I., Yin, X., Vose, R. S., Applequist, S., Arnfield, J. (2018), Enhancing the Data Coverage in the Integrated Global Radiosonde Archive, J. Atmos. Oceanic Technol. 35, 1753–1770; Durre, I., Vose, R. S., Wuertz, D. B. (2006), J. Climate 19, 53–68. NOAA NCEI, public domain.

Derived products (this repo's processed/) are released under CC-BY 4.0.

untuitivist — Fri, 14 Aug 2026 07:39:06 GMT

EgoMocap Dataset

Each row in metadata.jsonl describes one recording. New uploads use EgoMotionActions/<YYYYMMDD>/<primitive_id>/<session_id>/. The batch ID defaults to the uploader's local date when upload starts and may be selected explicitly. All pending sessions in one upload operation share that date, while upload_batch_created_at retains the exact ISO timestamp. A failed upload reuses its assigned date. Legacy YYYYMMDD_HHMMSS paths remain readable but are no longer generated. The built-in task catalog uses P01 through P29, but a safe custom single-directory name is also accepted.

Every newly prepared row contains numeric duration_s. For a full Session it is the sum of one Robocap reference-camera duration per included Segment, never the sum of all camera views. For an aligned-intersection upload it is the duration of the cropped common timeline. The same value is stored in that Session's manifest.json.

When the source directory follows mocap-<action:[A-Z]+<digits>>-S<collection-session-index>-<participant>-<count>p[<numeric-suffix>], both the dataset index and Session manifest contain a mocap_capture object with source_directory, action_id, collection_session_index, participant, and repetition_count. Generic mocap* names remain valid and omit this optional object. An optional numeric suffix after p is retained only in source_directory and does not change repetition_count. Action IDs accept one or more letters followed by one or more digits, and preserve leading zeros. Metadata-only corrections update the index and matching manifests in one commit without moving Session directories or uploading capture files again.

The dataset index and Session manifest also contain mocap_rigid_names: the actual exported entity names. objects contains deduplicated prop scan categories, for example ["BigBox", "Box"]. Trailing instance numbers are removed: Box0, Box1 and Box6 all map to Box. Scan bundles are stored under 3dobjs/; 3dobjs/index.json maps each category to its bundle. Scan categories use PascalCase without spaces, bracket lists or instance numbers (OpenBox.objcap). Original Mocap filenames and mocap_rigid_names remain unchanged. The mocap_rigid_names field retains all entity labels (for example Body0, Tracker0, BigBox0), deduplicated across formats. No objects or extra HH bodies are inferred. Backfills retain existing Session paths and remote-only files, merge manifests, and verify uploaded bytes. Upload preparation uses frame-count and timestamp reports; it does not automatically decode every video or run full source-content checks. Normal batches select matching frame counts; the separate fault uploader selects frame-count differences. Full structural/content inspection remains an independent operation. Fault metadata and manifests record quality_status=error, error_types and quality_details. Explicit full-inspection results may also include local_integrity and inspection_issues. Fault recordings with unmeasurable duration use duration_s=null, duration_status=unavailable and an explanatory duration_error.

At upload time, participants.jsonl is read from the destination dataset and matched against mocap_capture.participant (case-insensitive). Session index rows and manifests both receive participant_gender, participant_height_cm, and participant_weight_kg. Unavailable measurements remain JSON null. If the fault dataset has no catalog, the normal dataset catalog is used. Unmatched identities are reported; known remote values are retained only for the same participant. The catalog is read afresh for each upload batch and is never overwritten by Session uploads.

New ModelScope staging never applies lossy video compression. Full-session video files are copied byte-for-byte. Aligned-intersection video is encoded losslessly only when frame-accurate cropping requires it. Manifests that describe proxy-compressed video are rejected before upload.

EgoMotionActions/Demo/<primitive_id>/<session_id>/ contains recordings migrated from the legacy non-batched hierarchy. Demo is an example archive, not a new-upload destination.

Every session directory includes a self-contained timestamp_anomaly_detail_table.html inspection report. Paths inside manifests and reports are dataset-relative.

Capture sites

collection_site.id identifies the capture room: site_01 = 白墙场景 (white-wall room), site_02 = 绿幕场景 (green-screen room). Names are listed in collection_sites.json.

Download

Browse the ModelScope Dataset Files page for file metadata and individual data files.

:modelscope-code[]{type="sdk"}

:modelscope-code[]{type="git"}

Update a local copy

If the dataset was cloned with Git, run the following command inside the cloned repository to fast-forward it to the latest ModelScope revision:

git pull --ff-only

If Git reports that LFS objects are still missing after the pull, run git lfs pull. A directory downloaded with the ModelScope SDK is not necessarily a Git working tree; update that copy by running the SDK download command again with the same local directory.

After downloading, read the dataset-wide session index without any additional dependency:

import json
from pathlib import Path

dataset_root = Path(".")
records = [
    json.loads(line)
    for line in (dataset_root / "metadata.jsonl").read_text(encoding="utf-8").splitlines()
    if line.strip()
]
for record in records:
    print(
        record["primitive_id"],
        record["session_id"],
        record.get("duration_s"),
        record["session_path"],
    )

Global collection rules

Participant behavior: Move naturally; do not exaggerate gait or arm swing and do not imitate a robot.
Speed: Participants should not measure speed. Use the verbal descriptions: slow, normal, fast. Keep movement comfortable and controlled.
Start: Unless otherwise specified, begin standing naturally with feet approximately hip-width apart.
End: Unless otherwise specified, finish in stable standing and remain still for ~2 seconds.
Variation: For standardized trials, follow the listed geometry. For randomized trials, vary target positions, turn directions, path shape, and obstacle placement.

Action task catalog (P01-P29)

Each recording is assigned to exactly one action primitive. Episodes / participant is not yet specified in the collection definition and remains TBD for every primitive.

Speed & stopping

ID	Primitive	Participant instruction	Distance / geometry	Approx. duration	Room	Purpose
P01	Normal straight walking	Start still; walk straight at normal comfortable speed; stop at endpoint.	5 m	5–7 s	Open	Baseline forward locomotion
P02	Slow walking	Walk forward noticeably slower than normal; do not shuffle; stop.	5 m	7–10 s	Open	Slow-speed locomotion
P03	Fast walking	Walk quickly but do not run; remain comfortable and controlled; stop.	5 m	4–6 s	Open	Fast walking
P04	Accelerate	Start still; gradually go slow → normal → fast; briefly maintain fast speed; stop.	5 m	8–10 s	Open	Natural acceleration
P05	Decelerate	Start at normal speed; gradually slow over several steps; stop gently.	5 m	7–10 s	Open	Natural deceleration
P06	Sudden stop	Walk normally; at marked point stop as quickly as comfortably possible; remain still.	5 m	5–8 s	Open	Stopping dynamics
P07	Start from rest	Stand still 2 s; begin walking naturally; walk forward; stop.	5 m	7–9 s	Open	Stand-to-walk transition

Turning & trajectories

ID	Primitive	Participant instruction	Distance / geometry	Approx. duration	Room	Purpose
P08	Gentle curves	Walk forward; gradually curve left/right; straighten out.	~5 m path	7–10 s	Open	Continuous curvature
P09	45° turn	Walk straight; at marker turn ~45° while continuing to walk; continue straight.	~5 m	6–9 s	Open	Moderate heading change
P10	90° turn	Walk straight; turn 90° left/right while continuing to walk; continue.	~5 m	6–9 s	Open	Major heading change
P11	180° turn	Walk straight; turn around and walk back; use wide or tight turn.	~5 m	7–10 s	Open	Reverse heading
P12	Multiple turns	Walk through 3–5 turns in sequence; keep moving; stop at end.	~8–12 m	10–15 s	Open	Composed turning
P13	Figure eight	Walk one full figure-eight continuously at comfortable speed; stop.	~8–12 m	10–15 s	Open	Alternating curvature

Non-forward motion

ID	Primitive	Participant instruction	Distance / geometry	Approx. duration	Room	Purpose
P14	Sideways walking	Move laterally left/right for ~3 m; natural controlled steps; stop.	3 m	5–8 s	Open	Lateral locomotion
P15	Diagonal walking	Walk forward while moving diagonally left/right; finish ~45° from start direction.	~4–5 m	6–9 s	Open	Diagonal locomotion
P16	Backward walking	Walk backward slowly; stop.	3 m	6–10 s	Open	Backward locomotion

Composed & long sequences

ID	Primitive	Participant instruction	Distance / geometry	Approx. duration	Room	Purpose
P17	Speed change + turn	Walk normally; turn left/right while slowing; after turn accelerate back to normal.	~5–7 m	8–12 s	Open	Coupled speed/heading
P18	Stop → turn → walk	Walk; stop fully; turn 90° left/right while standing; walk new direction.	~5 m	8–11 s	Open	Discrete transition
P19	Walk → turn → stop	Walk; turn 90° left/right; walk several steps; stop.	~5 m	7–10 s	Open	Composed primitive
P20	S-shaped walking	Follow S-shaped path continuously; no stopping between curves.	~8–10 m	10–15 s	Open	Continuous heading changes
P21	Variable-speed natural walk	Walk continuously ~15 s; naturally alternate slow/normal/fast several times; no stop.	Room	15 s	Open	Natural speed trajectory
P22	Random locomotion	Walk through randomly placed floor targets; choose natural path; include turns and speed changes.	Room	15–25 s	Open	Goal-directed variation
P23	Long continuous locomotion	Walk continuously around room; naturally change direction/speed; no manipulation.	Room	30 s	Open	Long-horizon baseline

Full-body transitions

ID	Primitive	Participant instruction	Distance / geometry	Approx. duration	Room	Purpose
P24	Walk → crouch → walk	Walk; crouch comfortably at marker; remain briefly; stand and continue.	~5 m	10–15 s	Transition	Whole-body transition
P25	Walk → sit → stand → walk	Walk to chair; sit ~2 s; stand; walk away.	~5 m	12–18 s	Transition	Sit/stand transition

Terrain & navigation

ID	Primitive	Participant instruction	Distance / geometry	Approx. duration	Room	Purpose
P26	Step over obstacle	Walk toward low obstacle; step over without stopping; continue.	~5 m	7–10 s	Terrain	Foot placement
P27	Walk around obstacle	Walk toward obstacle; alter path to avoid it; continue toward original destination.	~5 m	7–12 s	Navigation	Obstacle avoidance
P28	Narrow passage	Walk through passage without touching sides; continue normally after exit.	0.7–1.2 m	7–12 s	Navigation	Body-aware navigation
P29	Target-directed navigation	Start at marker; visit targets in order; choose natural paths; avoid obstacles; stop at final target.	Room	15–30 s	Navigation	High-value goal-directed locomotion

Session contents

Required capture streams: six robocap_<segment>_video_*.mp4 first-person cameras, Robocap IMU/MAG databases, WAV audio per Segment, third-person videos, and Robowrist left/right video and sensor streams.
Standard motion-capture exports under mocap/: Body0 in TRC/BVH/CSV, Tracker0 (hat) in TRC/CSV, all three formats for each detected extra Body, and TRC/CSV for each named object. HH imposes no fixed extra-body count. Action-specific requirements can add exact object names. These export requirements are evaluated by independent full inspection, not the upload gate. Staging publishes only the explicitly selected files.
Optional artifact: RRD files, only when explicitly selected.
Generated: manifest.json and timestamp_anomaly_detail_table.html.

Raw calibration lives outside each action session under the dataset-level raw_calibration/<device_id>/ tree. Session manifests and metadata.jsonl contain only explicit main, left, and right device IDs needed to resolve those records.

Alignment variants

Session staging normally preserves the complete source recording. A session may instead contain only the frame-aligned intersection selected by its source Robocap-frame offset and Mocap/Robocap ratio. For an intersection session, manifest.json records the source half-open frame interval of every video and motion-capture file, the source offset, and the capture-time window applied to SQLite sensors. After cropping, all staged streams have zero residual offset. Original TRC/CSV/XRS frame and timestamp columns and the C3D first-frame identity remain unchanged for traceability.

Complete dataset structure

Additional native export formats and explicitly selected RRD files are optional. The standard Body0/Tracker0/object exports, capture streams and generated records are required for normal batches.

<dataset_root>/
  README.md                                      # required Dataset Card
  metadata.jsonl                                # required global session index
  raw_calibration/                              # required, maintained separately
    <device_id>/
  EgoMotionActions/                             # required action recordings
    <YYYYMMDD>/                                  # uploader-local date for new uploads
      <primitive_id>/                            # [A-Z]+<digits> convention or a custom name
        <session_id>/
          robocap_<segment>_video_*.mp4          # required six first-person videos
          robocap_<segment>_imu_*.db             # required Robocap IMU
          robocap_<segment>_mag_*.db             # required Robocap MAG
          audio_<segment>.wav                   # required WAV audio with actual samples
          mocap/
            *.mp4                                # required third-person video
            *Body0.{trc,bvh,csv}                 # required main body, all three formats
            *Tracker0.{trc,csv}                  # required hat, both formats
            *Body<N>.{trc,bvh,csv}               # extra tracked entities when present/configured
            *<object>.{trc,csv}                  # each named object, both formats
            *.{xrs,c3d,...}                      # optional additional exports
          robowrist_<device_id>_left/             # required left streams
          robowrist_<device_id>_right/            # required right streams
          rerun/<segment>/inspection/*.rrd        # optional
          manifest.json                          # generated, required
          timestamp_anomaly_detail_table.html    # generated, required
    Demo/                                        # legacy example recordings only
      <primitive_id>/
        <session_id>/                            # same session contents as above

README.md is the ModelScope Dataset Card. metadata.jsonl is the dataset-wide session index. raw_calibration/ is a required dataset-level collection maintained by the calibration workflow; session staging never copies session-local calibration files into it. Session preparation writes to local _prepared/<primitive_id>/<session_id>/; the upload operation assigns one shared local-time batch ID to new sessions and moves them into their final EgoMotionActions/ paths before uploading. Existing sessions keep their remote date/path during backfill. _prepared/ is never uploaded.

Before updating metadata.jsonl, the uploader downloads the current remote index and merges it with the local rows by (primitive_id, session_id). Updated fields replace matching values; unknown remote fields, unrelated batches and Demo records remain indexed. The combined manifest retains remote-only files, and both uploaded files and metadata are verified after transfer.

Cherrytest — Fri, 01 Aug 2025 00:57:07 GMT