Cross-stage experiments: code, paper figures (solve/exact/precision/recall)
Ubuntu+1: Cursor
a40741d
- _experiments/cross_stage/: per-cell prediction sweep across (method, train_stage, prompt_stage)
with analyze scripts that compute containment, difficulty-stratified solve
rate, failure-mode taxonomy, and a 3x3 cross-prompt grid.
- _runs/_paper_figures/: updated solve/exact and new precision/recall plots
with No-CoT-No-Curriculum horizontal baseline.
Co-authored-by: Cursor <cursoragent@cursor.com>
Resume scripts + paper figures
Avra98
624a68e
- Fix EVAL_RE regex bug in adaptive_k_cellpolicy_pipeline.py so plateau-based
k-bumping actually fires (metric was always None before).
- Add --initial_adapter so adaptive-k variants can warm-start from a saved
LoRA adapter directory.
- Make strawman_cellpolicy_pipeline.sh accept an optional SFT_INIT env var
for warm-starting from a previously trained baseline checkpoint.
- New launch_finish_repos23.sh: brings up the 6 strawman warm-start
variants and 2 adaptive-k resume variants across all 8 GPUs.
- Add paper figures comparing latent vs vanilla baseline solve / per-cell
exact across the three curriculum stages (plot_stage_progression.py +
stage_progression_{solve,exact}.{pdf,png}).
- Update .gitignore to keep in-flight run output trees out of the code repo.
Loosen .gitignore on the HF repo (keep data/ + *.jsonl)
