mirror of
https://github.com/ranausmanai/tinyforge-zero.git
synced 2026-06-08 20:55:13 +02:00
6305ff0f91
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)
13 lines
228 B
Text
13 lines
228 B
Text
# Exact stack used in the paper (May 2026).
|
|
# Tested on H100 80GB (driver 570+) and RTX 6000 Ada.
|
|
|
|
torch==2.6.0
|
|
transformers==4.51.3
|
|
vllm==0.8.5
|
|
peft==0.13.0
|
|
accelerate>=0.30
|
|
datasets>=2.18
|
|
sympy>=1.12
|
|
numpy<2
|
|
scipy
|
|
hf_transfer
|