weightpatch-run-math-gsm8k / trace /final_answer.md
Perfect7613's picture
WeightPatch run astra-math-1
1bde957 verified
|
Raw History Blame Contribute Delete
931 Bytes

No edit produced a reliable gain. All changes were reverted.

  • Diagnosis: On the initial 128 problems, 27 answers were wrong: 6 truncated and 21 finished incorrectly. Most failures were reasoning or arithmetic errors.
  • LaTeX hypothesis: Weight rankings, activation contrast, and traces implicated L33N10488. Halving its output column fixed 0 answers and broke 1; zeroing it fixed 1 and broke 1.
  • Verbose-introduction hypothesis: Two localizer rankings and traces implicated L34N5892. Reducing its output column by 25% fixed 1 answer and broke 2.

Neither experiment exceeded statistical noise.

Full dev accuracy: 196/256 (76.56%) before and after restoration. Harmful refusals remained 36/48; benign refusals remained 1/48, with identical per-prompt outcomes.

Grader: Not submitted—no qualifying gain. Stopped after two unsuccessful hypothesis rounds. Final weights are unchanged.