Files
stack/tests
kert ad9ee29624
Some checks failed
CI / lint (push) Successful in 1m36s
CI / notebooks-smoke (push) Successful in 2m6s
Deploy / changes (push) Successful in 18s
CI / test (push) Failing after 5m36s
Deploy / zotero (push) Has been skipped
Deploy / docs (push) Successful in 2m24s
Deploy / notebooks (push) Successful in 6m4s
Deploy / api (push) Successful in 2m13s
Deploy / llm (push) Successful in 1m38s
Deploy / mc (push) Has been skipped
Deploy / bcda (push) Successful in 2m20s
Deploy / report (push) Successful in 15s
feat(llm): tagger shortlists 12 themes with a 0.45 floor
Golden v2 (200 comments, 8 dockets) showed recall, not precision, was
the problem: stored tags scored P 0.96 R 0.55 on the held-out half.
Live tune-split sweep on the rig: 14b/top 8/0.60 F1 0.72; /0.50 0.76;
top 12/0.45 0.78; qwen2.5:32b was more conservative (top 8/0.50: P 0.97
R 0.59 F1 0.73). Held-out with top 12/0.45: P 0.95 R 0.61 F1 0.74 —
PFS dockets 0.67-0.91, IPPS/OPPS ~0.28 (topics outside vocab v2).
2026-11-11 15:40:12 -05:00
..
2026-10-31 22:00:04 -04:00