Reorganizes the repo so every section of the paper has a corresponding
script. Previously only the core recipe + control + evals were here.
New subdirs:
- tts/ — test-time sampling (§2.2, §3.3): scaling sweep, HE, MATH-500,
AIME, 14B-recipe + TTS, 8B-raw-TTS control.
- experiments/ — every §3 finding as a runnable script:
· self_consistency (§3.4)
· recipe_x_tts_synergy (§3.5, novel)
· mbpp_seeded_cross_arch (§3.9)
· cross_domain_code_to_math (§3.10)
· self_correction_math_{naive,fixed} (§3.10, the
catastrophic-then-recovered case)
· math500_seeded_mining (§3.10 distribution mismatch)
· bcb_hard_eval (§3.10 distribution mismatch)
· recursive_bootstrap (§3.10 plateau)
· diversity_cued_mining (§3.10 low yield)
· aime_scaling (TTS curve)
· star_baseline_gsm8k (related-work baseline)
- evals/ — moved out of recipe/ (eval_raw, eval_plus, confirm)
Also adds: bootstrap_14b_4bit_harvest, curriculum_code, math_bootstrap to
recipe/ for completeness.
REPRODUCE.md now maps each paper section / table / figure to its exact
script and expected output.
Companion artifact for the paper 'How Far Can an Open Base Model
Self-Improve? Recipes, Limits, and Test-Time Synergy'.
Contents:
- recipe/{train_on_pairs,bootstrap,multi_pair_14b,curriculum_math,eval_raw,eval_plus,confirm}.py
- data/pairs_{7b_40,14b_multi_new60,math_13}.jsonl (released mined pairs)
- controls/mbpp_corrupt_control.py (the +0 negative control)
- docs/{scaling_chart,fig1_headline,fig6_boundary}.png
- REPRODUCE.md (paper claim -> exact command mapping)
