Downloads
- Discoveries: discoveries.json (full records, including evidence, optional review scope, follow-up and relationships) and discoveries.csv (evidence and caveat columns; structured review, follow-up and relationship fields encoded as JSON cells).
- AI-found vulnerabilities: vulnerabilities.json and vulnerabilities.csv (CVE identifiers separated by semicolons, with classification and remediation caveats). An unknown
ai_attribution_dateisnullin JSON and an empty cell in CSV; consumers must handle this separately from a dated account. These cases are separate from the scientific count. - Model progress: model-progress.json, the separately imported capability and benchmark snapshot from Epoch AI, CC BY 4.0. Import and interpretation. Source entries can overlap notable releases and must not be added to that collection’s count.
- SWE-bench Verified: swe-bench.json, a separate selection of official mini-SWE-agent evaluation runs, under CC BY-NC 4.0 (noncommercial reuse only). This component is excluded from the catalog’s CC BY 4.0 grant and current Hugging Face catalog exports. Scope and attribution.
- Additional benchmark evidence: 10-component directory, with per-source attributed JSON downloads and distinct CC BY 4.0, CC BY-SA 4.0, MIT or Apache-2.0 terms. These snapshots are excluded from the blanket catalog license and its Hugging Face export. Selections, attribution and reuse rules.
- Models: models.json, generated from Epoch AI data, which keeps its own CC BY 4.0 attribution.
- Every file is also available from the repository, whose releases are archived on Zenodo.
CSV cells beginning with a spreadsheet formula character are prefixed with an apostrophe; JSON retains the original values.
Static API
There is no server: the API is a set of static JSON files, regenerated with the site, so responses are cached by GitHub Pages for about ten minutes and allow cross-origin reads. Filtering is done by the client.
- api/v1/index.json: counts, a SHA-256 fingerprint of the active discoveries, vulnerabilities, models and archived discoveries, the latest discovery addition date, license, citation and endpoint addresses. The legacy
data_version_dateis an alias for that addition date, not a last-modified date. api/v1/discoveries/ID.jsonandapi/v1/vulnerabilities/ID.json: one record per file. Example: api/v1/discoveries/2020-02-halicin.json.- The complete active lists are the JSON files under Downloads. IDs remain stable, including historical slugs, but archived discoveries are removed from the active API; their historical share pages remain available.
- The version in the path changes only if a field is removed or renamed.
The fingerprint hashes UTF-8 JSON containing those four datasets in that order, with sorted object keys, compact separators and original array order. It identifies content, not a review date.
Schemas
Records follow JSON Schema (draft 2020-12): discoveries.schema.json and vulnerabilities.schema.json. Field meanings are described in the repository documents docs/discovery-schema.md and docs/vulnerability-schema.md. An empty paper_url means no paper link is established, not that none exists.
Counts describe catalog entries, which can overlap or bundle related results, rather than independent findings. See evidence annotation coverage and primary source composition by year before interpreting missing annotations or comparing subsets.
Examples
- Shell:
curl -s https://aisciencetracker.org/api/v1/index.json - Python:
pandas.read_csv("https://aisciencetracker.org/data/discoveries.csv") - JavaScript:
fetch("https://aisciencetracker.org/data/discoveries.json").then(r => r.json())
License and citation
Catalog content is licensed under CC BY 4.0: credit “AI Science Tracker” with a link. Cite the archived catalog as doi:10.5281/zenodo.24217686; this cites the catalog, not the underlying scientific papers, which you should cite from each record’s source. Figures are reported by sources and have not been independently verified, and coverage is manual and incomplete. See the methodology for what is included.
Changes
The changelog lists data and format changes, and the Atom feed lists discovery additions by addition date and vulnerability cases by source-check date; it is not a complete change history.