This folder contains the v3.1 predictive-hardness reweighting prototype.
Archived as a negative result. The implementation worked, but the main
soft_hard strategy underperformed the uniform baseline on both recorded seeds.
Final interpretation:
- Predictive hardness can be computed and connected to training.
- Chunk-level reweighting alone was not enough to improve control.
- This path motivated later loss-level or representation-level supervision ideas.
| File | Purpose |
|---|---|
config.py |
Shared PHRew config |
dataset.py |
Dataset utilities |
chunk_scorer.py |
Chunk hardness scoring |
stage2b_score_chunks.py |
Score chunks with the trained probe |
stage3_train_phrew.py |
Train PHRew variants |
stage4_generate_cmds.py |
Generate comparison commands |
See ../Paper/研究方向v3.1_预测难度重加权.md
for the full reasoning and results.