Scripts and results for “Removing a loop direction from a small model’s weights”

These are the scripts behind the post, as they ran, with three kinds of edit for publication: four comment lines (run-32k.sh, stage3.sh, stage4.sh, sweep.sh) lost references to our chat, provenance.sh no longer prints the machine’s hostname, and PROVENANCE.md has its hostname and one home-directory path removed. Nothing that computes a result was changed. The sha256 hashes in PROVENANCE.md are those of the original files, so those five scripts hash differently here.

• collect.py: base model answers MATH train problems (40 per subject) via llama-server.
• acts.py: finds where each answer starts looping and averages the residual stream before and inside the loop.
• rank.py: prints, per layer, how much the MLP output points along the loop direction.
• edit.py (edit A) and edit2.py (edit B and the sw-* variants): the down_proj column edit.
• evalmath.py: MATH-500 evaluation (math_verify), one record per problem.
• run-editB.sh, run-ref-p50c.sh, run-32k.sh, sweep.sh, stage3.sh, stage4.sh: the runs, in the order they happened.
• provenance.sh and PROVENANCE.md: software versions, file hashes, and every run’s summary line.
• base-h400-32k.jsonl, editB-h400-32k.jsonl, ref-p50c-h400-32k.jsonl, sw-a4k256-h400-32k.jsonl: per-problem results on MATH-500 problems 100-499 (base, our edit B, Astra’s edit rebuilt as ref-p50c, and the sw-a4k256 variant). Pair them by “idx”.

Astra’s reference edit is not copied here. It comes from https://github.com/RikiyaT/lm_understand-logs at commit 4d09ba6d6a7e891dd8063974dfe81b3d78e3494f (folder exp8_math; run-ref-p50c.sh expects its code in ref/code and the p50c run in ref/runs).