Skip to content

datasets

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

datasets

README.md

Community-contributed datasets supporting research, benchmarking, and practical security testing for GenAI data security. Each subfolder contains a specific dataset aligned with the DSGAI risk taxonomy.

Folder Dataset Status
vulnerability_dataset Real-world vulnerabilities affecting LLM applications Active
exploit_dataset Documented exploits and attack techniques targeting LLMs Active
riskassessment_dataset Mapped risk assessments for various LLM deployments Active
incident_dataset Anonymized GenAI data security incidents (DSGAI01–DSGAI21) Accepting contributions
promptinj_dataextraction_testcases Adversarial prompts and extraction techniques for red-teaming Accepting contributions
rag_dataset RAG poisoning and retrieval integrity test sets Accepting contributions
crossframework_mapping_dataset Machine-readable DSGAI-to-framework control mappings Superseded by the GenAI Crosswalk
agentdataflow_toolexchange_traces Sanitized agentic AI tool call and delegation chain traces Accepting contributions

Metadata, releases and citation

Croissant. Every published dataset folder has a croissant.json: MLCommons Croissant 1.1 metadata with Responsible AI fields (limitations, biases, intended uses, personal data, collection, maintenance). ML tooling can load the records from it directly:

import mlcroissant as mlc
records = mlc.Dataset(jsonld="datasets/vulnerability_dataset/croissant.json").records(record_set="index")

The files are generated by data_validation/croissant/build_croissant.py, which also holds each dataset's RAI text. CI fails when one is out of date, for example after an entry changes the dataset's index.csv, so regenerate after editing entries.

Releases. Versioned releases are cut by tagging datasets-v<version>, where the version must equal the one in CITATION.cff. The release workflow re-runs every check and publishes one zip per dataset plus an all-datasets zip, with SHA256SUMS and a Sigstore build-provenance attestation for each file. To verify a download:

sha256sum -c SHA256SUMS --ignore-missing
gh attestation verify owasp-dsgai-vulnerability_dataset-1.0.0.zip --repo GenAI-Security-Project/GenAI-Data-Security-Initiative

Citation. Cite the datasets with CITATION.cff (GitHub's "Cite this repository" button), or with the per-dataset citeAs in each croissant.json. With the repository enabled in Zenodo's GitHub integration, each release is archived on Zenodo with a DOI; .zenodo.json supplies its metadata.

How to Contribute

  1. Fork this repository
  2. Add your data to the appropriate dataset folder following the formatting guidelines in that folder's README
  3. Run python data_validation/run_all_checks.py, and regenerate the dataset's index.csv (where it has a build_index.py) and croissant.json (python data_validation/croissant/build_croissant.py, which needs pip install -r data_validation/requirements-croissant.txt)
  4. Submit a pull request with a description of your contribution

For questions or to discuss contributions, join #team-genai-data-security-initiative on the OWASP Slack workspace.