Skip to content

Run a prospective behaviour check on 259 advisories published after the snapshot - #62

Merged
professor3333 merged 1 commit into
mainfrom
eval/prospective-check
Nov 15, 2026
Merged

professor3333 merged 1 commit into
mainfrom
eval/prospective-check

Conversation

@professor3333

Copy link
Copy Markdown
Owner

Prospective check (Definition of Done: prospective batch reported separately), behaviour only: no labels exist or will be made, so no accuracy or agreement.

  • scripts/experiments/prospective.sh [BUNDLE]: fetches github/advisory-database at 9b0169eb into its own directory (pinned snapshot untouched), builds it with the same pipeline (reports/prospective/snapshot-9b0169eb.json), freezes prospective-1.json (259 PyPI advisories absent from the pinned corpus and published after 2026-10-20T03:38:30Z), predicts with the frozen CRF (live review policy) and the run-3 DeBERTa, and writes summary.md against the 272 dev documents.
  • Findings: median 515 words vs 79 on dev; 71% exceed one 512-token window (dev 13%); 29% exceed 2,000 tokens (dev 3%); advisory length has risen through 2026 (2025-H2 133 → 2026-H1 215 → 2026-H2 394 → prospective 515). CRF: 4.37 conditions/doc (dev 1.11), "no explicit conditions found" 7.7% (dev 43.8%), negation flag on 21% of conditions (dev 8%). DeBERTa: 15.3 conditions/doc (dev 4.5).
  • README (phase 7 row + paragraph) and dataset card (known biases) updated.

No test data touched.

…he snapshot

The advisory database at 9b0169eb (2026-11-15) is built into separate paths; 259 PyPI advisories absent from the pinned corpus and published after its commit date form the frozen prospective-1 batch. There are no labels, so only behaviour is reported, against the 272 dev documents of batch-03, batch-06 and batch-07.

New advisories are much longer (median 515 words against 79; 71% need more than one 512-token window against 13%), continuing a rise through 2026. The CRF suggests 4.4 conditions per advisory (dev 1.1) and abstains on 7.7% (dev 43.8%); DeBERTa suggests 15.3 (dev 4.5). NOT_REQUIRED, EXCEPTION and negation-flagged items are more common.
@professor3333
professor3333 merged commit 3063851 into main Nov 15, 2026
2 checks passed
@professor3333
professor3333 deleted the eval/prospective-check branch November 15, 2026 03:23
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant