kev-catalog / README.md
rpotham's picture
KEV catalog v1 (full splits): 56 benchmark / 114 train / 182 code_fix
dd7dc7d verified
|
Raw History Blame Contribute Delete
2.75 kB
---
pretty_name: CISA KEV catalog (enriched)
license: other
tags:
- security
- vulnerability
- kev
- cisa
configs:
- config_name: raw
data_files:
- split: full
path: raw/full.jsonl
- config_name: code_fix
data_files:
- split: full
path: code_fix/full.jsonl
- config_name: prioritize
data_files:
- split: full
path: prioritize/full.jsonl
- config_name: benchmark
data_files:
- split: test
path: benchmark/test.jsonl
- config_name: train
data_files:
- split: train
path: train/train.jsonl
---
# CISA KEV catalog → discovery / prioritization assets
Enriched CISA Known Exploited Vulnerabilities (KEV) rows with free OSV
`code_fix` where available. Splits follow the experiment
[`SPLIT_CONTRACT`](https://github.com/patcherai/vms/blob/ram/experiments/experiments/103026-kev-catalog-hf-benchmark/SPLIT_CONTRACT.md).
Split name `full` is used instead of `all` because Hugging Face `datasets`
reserves `all` as the union-of-splits keyword.
## Provenance
| Field | Value |
| --- | --- |
| KEV catalogVersion | `2026.10.30` |
| KEV sha256 | `3aa9cf3c01f116362a29d111702979175453c425e3733bde80fc4f6f8e2294e5` |
| OSV enrich observed_at | `2026-10-31T01:05:09.836782+00:00` |
| Split freeze content hash | `8286776291c9ab34b278552b583b9a244d49bbb39d5b223f1c35c6f84687f678` |
| Source experiment | `experiments/103026-kev-catalog-hf-benchmark` |
## Configs
| Config | Split | Rows | Purpose |
| --- | --- | ---: | --- |
| `raw` | `full` | 1766 | All KEV CVEs + enrichment status (incl. non-fix) |
| `code_fix` | `full` | 182 | KEV ∩ OSV with ≥1 non-apk/image code_fix |
| `prioritize` | `full` | 1766 | All KEV + ransomware/dates (EPSS null in v1) |
| `benchmark` | `test` | 56 | Held-out discovery tasks |
| `train` | `train` | 114 | Remaining code_fix after exclusions |
Hugging Face `datasets` reserves the split name `all`, so raw / code_fix /
prioritize use split `full` (SPLIT_CONTRACT intent unchanged).
## Leakage checklist (v1 freeze)
- Advisory-ID overlap train ∩ Golden-30 / DF-Bench / low_level_hard20: **0**
- Repo+commit overlap train ∩ those evals: **0**
- `benchmark` ∩ `train` at CVE level: **0**
- code_fix rows hard-excluded from both splits (eval identity): **0**
- code_fix excluded from train due to luna-not-both / patch-data cross-link: **12**
Non-`code_fix` statuses remain in `raw` / `prioritize` with `status` and `error`
metadata (`osv_http_error`, `osv_no_code_fix`, `osv_fetch_error`).
## License / upstream
CISA KEV is a U.S. government work. OSV records follow their upstream licenses.
This packaging does not grant rights to private or non-public repositories that
may appear as `code_fix` URLs.