MIT AI Risk index (bundled data)¶
FairSense ships a FAISS index built from the AI Risk Database sheet of the MIT AI Risk Repository (CC BY 4.0).
| Field | Bundled index (v0.3.0 and later on this release line) |
|---|---|
| Repository version | V4 |
| Export date | 3 December 2025 |
| Indexed risks | 1,538 (Risk Category / Risk Sub-Category rows) |
| Embeddings | all-MiniLM-L6-v2, 384 dimensions |
Constants: fairsense_agentix.data.risk_corpus.
The JOSS paper and Zenodo v0.2.2 archive
describe the earlier V3 snapshot (1,340 risks, 26 March 2025). Rebuild from
source with scripts/transform_mit_data.py --repository-version v3 to reproduce that
corpus.
RMF (mitigation) rows¶
Bundled ai_rmf.csv / rmf_meta.json actions are illustrative NIST AI RMF–style
placeholders produced by scripts/transform_mit_data.py (domain → GOVERN / MAP /
MEASURE / MANAGE templates). They are linked by risk_id to each indexed MIT risk
for structured display—they are not direct quotes from the MIT spreadsheet and
should not be cited as MIT repository mitigations.
To ship MIT-sourced measures instead, extend the transform to read mitigation columns from a future export (when present) and rebuild indexes.
How RMF rows are attached at runtime¶
After FAISS returns top-k MIT risks, search_rmf_per_risk loads mitigations from
rmf_meta.json via fairsense_agentix.graphs.risk.rmf_lookup—keyed by MIT
risk_id, not by a second global FAISS query on the RMF index. That avoids
duplicate actions when many risks share similar template embeddings. Request
options["rmf_per_risk"] (default 4, max 20) to cap how many actions appear per
risk; the UI uses the same default for risk analysis.
Rebuild from a new MIT export¶
See data/raw/README.md in the repository.