NeurIPS 2026 Evaluations & Datasets Hosting Guidelines

Please note: These guidelines will be updated on a rolling basis until April 15, 2026 (OpenReview portal opens for submission). Authors should revisit this page regularly to avoid misunderstandings or missing important updates.

This webpage provides guidance for authors submitting to the Evaluations and Datasets Track. Authors of datasets are **required** to make their datasets available along with Croissant machine-readable metadata (core and minimal RAI) to streamline the review process and meet industry standards for high-quality documentation, reproducibility, and accessibility. 

The dataset hosting process is outlined below:

1. Host your dataset. Choose one of the supported platforms: Harvard Dataverse, Kaggle, Hugging Face, and OpenML or set up a custom hosting solution. The goal is to ensure that the dataset is properly hosted, also after the conference.
2. Create the Croissant file - core fields. Use the platform’s tooling to download the automatically generated Croissant file, or create it manually. This file will initially include only the Croissant core fields.
3. Complete the Croissant file - RAI fields. Add the required Responsible AI (RAI) metadata. We provide additional tooling for this, see below.
4. Validate the Croissant file. Use the online Croissant validator tool (see below)  to ensure your file is complete and correctly formatted.
5. Provide dataset URL. Include a link to your hosted dataset and upload the completed Croissant file to OpenReview.
6. Post-acceptance requirement. If your submission is accepted, you must make your dataset publicly available by the camera-ready deadline.

 

1. Host your dataset

Harvard Dataverse, Kaggle, Hugging Face, and OpenML are the preferred platforms for hosting datasets. These platforms automatically generate a Croissant file and enable programmatic metadata verification and dataset assessment, thereby streamlining and standardizing the review process. When you make your dataset accessible via one of these platforms, submitting will be as simple as providing a dataset URL and uploading the generated Croissant file.

The table below outlines key platform features to help authors choose where to host their dataset. Authors may make their dataset accessible via multiple platforms at any time.

 

	Automatically generated Croissant file	API: Croissant download, data download, data loader	Hosting restrictions	Private preview URL access	Credentialized (gated) access	Supports Linking to Papers	DOIs
Harvard Dataverse	✅	✅✅✅	1TB per dataset (2.5GB per file)   Any file types	✅	✅	✅	✅
Kaggle	✅	✅✅✅	200GB per dataset   Any file types	✅	❌	❌	✅
Hugging Face	✅	✅✅✅	300GB per dataset public   Any file types	❌	✅	✅	✅
OpenML	✅	✅✅✅	200GB per dataset Any file types	❌	❌	✅	❌

Authors are responsible for reviewing and complying with the Terms of Service of the platform(s) they choose to use.

If you need more details on how to publish your dataset on these platforms, please refer to the “How to publish on preferred hosting platforms” section at the end of this page.

Note on Croissant file: Please note that these platforms populate only the “core” Croissant fields and you need to ensure that you add the minimal Responsible AI as described below. 

 

Self-hosting your dataset and other data storage platforms

If you choose NOT to release your dataset via one of these preferred platforms, you can self-host the data or use other platforms, but you will still be required to make your dataset accessible to reviewers via URL (e.g., to a GitHub repo, public cloud bucket, Zenodo, etc.) and manually generate and upload a Croissant file representation of your dataset as part of the OpenReview submission process.

 

 

2. Core Croissant metadata

As part of the submission process, we require submissions with dataset contributions to provide a Croissant file with both Core and (minimal) Responsible AI metadata fields. 

If you hosted your dataset in one of the platforms listed above, they will generate the Croissant file for you. You can download it directly from the platform and jump to the next step. Step-by-step guidelines on downloading the Croissant file are in the table at the very bottom of this guide. If you use other or bespoke forms of dataset hosting, create the Croissant file as explained next.

 

Creating Croissant files manually

A Croissant file is a JSON file that describes your dataset in enough detail to load and analyse it automatically. For more information, check the Croissant specification. 

If you prefer to generate your Croissant metadata locally, you can create the JSON file manually, use the Croissant Python package, or use the Croissant Baker CLI tool (currently in alpha), which automatically analyzes your dataset folder and generates a Croissant file. 

 

Minimal Croissant core metadata

NeurIPS follows the standard Croissant spec, meaning that these fields are required:

 

 

3. Responsible AI metadata

It is important that datasets are created and used responsibly. Authors must provide a minimal set of Responsible AI (RAI) metadata, as described below. This information must be clearly identifiable in the submission, and reviewers will check for its presence.

The RAI metadata must be included in the Croissant file, which serves as the canonical, long-term record accompanying the dataset. Wherever possible, authors should provide the full RAI information directly in the corresponding Croissant fields.

If certain information cannot be fully expressed in Croissant (e.g., due to formatting constraints such as LaTeX equations), authors may instead use the Croissant fields to clearly point to the relevant sections of the paper (main text or appendix).

Authors may also include the RAI information in other parts of the paper (e.g., main paper or appendix) for clarity or elaboration. In all cases, we ask that the authors at least touch on these topics at a high level in the main part of the paper, even if only to guide the reader to where more detailed information can be found (e.g. appendix or Croissant).

All submissions that include a dataset as a contribution must provide answers to these RAI fields, at least in the Croissant file. Failure to do so may constitute grounds for rejection.

 

Minimal Responsible AI metadata

The table below contains a high-level overview of the minimal RAI data. A much more detailed description with examples can be found in the NeurIPS Evaluations and Datasets RAI guideline.

		Data limitations  rai:dataLimitations		Known constraints on the dataset's applicability: distributional gaps, underrepresented populations, data quality issues, or domain restrictions. Also list any uses for which this dataset is explicitly not recommended.
		Data biases  rai:dataBiases,		Any known or suspected biases in the data, including selection bias, label bias, or demographic skew. Which population groups or scenarios may be over- or under-represented, and how this may affect model behaviour.
		Personal or sensitive information rai:personalSensitiveInformation		Personal or sensitive information such as Gender, Socio-economic status, Geography, Language, Age, Culture, Experience or Seniority, Health or medical data, Political or religious beliefs.
		Data use cases rai:dataUseCases		Construct validity: what real-world concept the data is intended to measure or represent, and provide evidence that the data reliably captures that construct. List the use cases for which validity has been established (e.g. safety evaluation, fairness auditing, fine-tuning), and for which it has not.
		Social impact rai:dataSocialImpact		The potential positive and negative societal effects of using this dataset, including risks of misuse, fairness implications for specific communities, and any mitigations put in place.
		Synthetic data rai:hasSyntheticData		A boolean indicating the presence of synthetic data. If so, authors are encouraged to describe the synthetic data process in using the data collection and annotation fields.
		Source datasets prov:wasDerivedFrom		The URI of the dataset from which the present dataset is derived. Can be many URIs from different data sources. For synthetic data point to synthetic data seeds used.
		Provenance activities prov:wasGeneratedBy		Preprocessing: cleaning, or filtering steps applied to the data prior to use, to make the data pipeline reproducible. Data collection: collection period, geographic scope, and any instruments or protocols used. In case of synthetic data, document the seeds or prompts used or the synthetic data generator used. Data annotation: labeling schema, instructions provided to annotators, quality control measures, and inter-annotator agreement scores where available. Describe the human teams, synthetic agents and platforms used to collect, annotate or preprocess the data.

Including RAI metadata in the Croissant file

The hosting platforms above currently only populate the “core” Croissant fields, so you need to ensure that you add the RAI fields. There are several options to do this:

Croissant RAI tool
We will provide an online tool (coming soon) where you can easily augment your existing Croissant file with the minimal RAI metadata.

Manually add the RAI fields
Adapt the Croissant file by adding the fields specified in the table and guidelines above. Afterwards, use the Croissant validator (see the next step) to check whether it was done correctly.

 

4. Validate the Croissant file 

When you submit to Open Review, you will be required to provide a URL to your dataset accessible to reviewers and a Croissant metadata file. To avoid desk rejections, we ask authors to verify the validity of their Croissant file using the Croissant validator tool. 

Please use this online tool to verify your Croissant file. This tool will check whether it is valid, sufficiently complete, and whether the data can be automatically accessed and loaded (the latter may not be possible for all datasets). The same information will be made available to reviewers. 

Note: this tool may experience heavy usage near the submission deadline. Please run the validation well in advance. On the bottom of the tool website, you can also find workarounds to duplicate and run the tool yourself.

 

5. Submission

In your submission on OpenReview, you must always include the following:

1. The URL to your dataset *accessible* to reviewers
2. The Croissant metadata file (validated)

After submissions are closed, if your Croissant file is invalid or if your data is not accessible, your submission may be desk-rejected. Otherwise, reviewers will begin reviewing your submission, paper, dataset, and metadata. 

If your dataset is accepted, you will be required to make it public by the camera-ready deadline. Failure to do so may result in removal from the conference and proceedings.

 

 

Appendix: Preferred Hosting Platforms

This section provides specific guidance and documentation on how to make your dataset and its Croissant metadata file accessible via the preferred hosting platforms (Harvard Dataverse, Kaggle, Hugging Face, and OpenML) and how to download your Croissant file.

 

	How to upload	How to download (files, Croissant)	How to get help, e.g., to request additional storage quota
Harvard Dataverse	Upload a Dataset via UI (after login) or CLI)   Requirements An email-verified Harvard Dataverse user account Publicly shared (or “Link Sharing” turned on in the “Edit” menu  to generate a Preview URL for a private dataset) at time of submission to E&D track   Platform restrictions Documentation 1TB per dataset (2.5GB per file)	Download files Click “Access Dataset” and choose an option   Download Croissant Click “Metadata”, Click “Export Metadata”, and select “Croissant” Via Python client	Contact support@dataverse.harvard.edu
Kaggle	Upload a Dataset via UI (after login) or Python client   Requirements A phone-verified Kaggle user account Publicly shared (or “Link Sharing” turned on in “Settings”  to generate a Preview URL for a private dataset) at time of submission to E&D track (Optional) An approved Organization profile to host the data under if preferred, e.g., your research lab   Platform restrictions Documentation 200GB per dataset	Download files Click “Download” and choose “Download dataset as zip” Via kagglehub Python client   Download Croissant Click “Download” and choose “Export metadata as Croissant” Click “Code”, select “Load via mlcroissant”, and copy Python code	Contact kaggle-datasets@kaggle.com
Hugging Face	Upload a Dataset via UI (after login) or Python client   Requirements A Hugging Face user account Publicly shared at time of submission to E&D track Must be a format listed here in order to generate a Croissant file (Optional) An Organization profile to host the data under if preferred, e.g., your research lab   Platform restrictions Documentation 300GB per dataset	Download files Click “API” and copy curl code   Download Croissant Click “Croissant” and choose “Download Croissant metadata” Via Python client Download files Click “API” and copy curl code   Download Croissant Click “Croissant” and choose “Download Croissant metadata” Via Python client	Contact datasets@huggingface.co
OpenML	Upload Dataset via UI (after login) or Python client (recommended)   Requirements An email-verified OpenML user account Publicly shared at time of submission to E&D track   Platform restrictions Documentation 200GB per dataset	Download files Click “Download” Via Python client   Download Croissant In the web UI, click “Croissant”.	Contact openmlhq@openml.org

 

 

FAQ

Please check the Evaluations and Datasets Track FAQs