latkes/inside-out-replication-v2-pamqfix-canary-v1
Inside-Out Replication V2 — P(a|q) A.8.1-fix CANARY Corrected externals (04_external_scores.py with the A.8.1 in-context tokenization fix) over a deterministic 25-question-id subset of Mistral-7B-Instruct-v0.3 / P264 / test (5832 candidate (q,a) rows). Bounded subset used to validate the fix end-to-end on MLL before the full 3×4 re-score (per preflight rev2 spec; see red_team_brief.md). Canary self-validation C2 rows scored: 5832 (no build_qa_sequence ValueError;… See the full description on the dataset page: https://huggingface.co/datasets/latkes/inside-out-replication-v2-pamqfix-canary-v1.
Inside-Out Replication V2 — P(a|q) A.8.1-fix CANARY
Corrected externals (04_external_scores.py with the A.8.1 in-context tokenization fix) over a deterministic 25-question-id subset of Mistral-7B-Instruct-v0.3 / P264 / test (5832 candidate (q,a) rows). Bounded subset used to validate the fix end-to-end on MLL before the full 3×4 re-score (per preflight rev2 spec; see red_team_brief.md).
Canary self-validation
- C2 rows scored: 5832 (no
build_qa_sequenceValueError; would have crashed the job otherwise). - C5 rows failing eot ≤ content / norm ≤ 0 / NaN: 0 / 5832.
- C1 rows whose
log_p_a_qchanged vs the OLD buggy-concat phase4 (same (qid,answer) keys): 360 / 5832 (6.2%).
Subset K vs committed full-pool corrected baseline (Mistral P264)
Note: subset K (25 qids) vs baseline K (442 qids) are NOT directly comparable — magnitudes drift with subset choice. The apples-to-apples check is C1 above (per-row diff on identical keys).
Probe is DEFERRED in this phase (see NEXT_STEPS_TODO.md).
Reproducibility
The fix is on branch fix/judge-heuristic of the project repo (commits 60dc183 → 6a12f03). The pre-flight + canary harness lives in sbatch/pamqfix/. The no-GPU correctness gate is tests/test_qa_sequence.py.
