Scale RL features
Open
algorithm
enhancement
- Dominant language
- Python
- Stars
- 2k
- Forks
- 561
- Avg merge
- 4d 5h
- Merged PRs (30d)
- 145
Description
- [ ] advantage normalization (batch level)
- [ ] zero-variance filtering
- [ ] FP32 lm head (train and gen)
- [ ] CISPO
- [ ] no-positive resampling (data curriculum: skip >=0.9 pass rate)
cc/ @parthchadha
Contributor guide
Assessment
This issue has not been assessed yet.