chi2liu / ABC-GRPO
Code For All-Quadrant Bounded Clipping GRPO. arxiv.org/pdf/2601.03895
Why this job: topics: rlhf, reinforcement-learning
What it does
- 4-Boundary Clipping. Independent epsilon values (ε₁, ε₂, ε₃, ε₄) for all four (r, Â) quadrants
- Unconditional Bounds. Prevents unbounded gradient updates regardless of advantage sign
- Entropy Preservation. Maintains substantially higher entropy (10.9× higher than GRPO) throughout training
- Improved Generalization. Achieves monotonic improvement in Pass@64 while GRPO degrades
- Q4 Protection. Addresses the critical blind spot where 41.4% of clipping events occur
Get started
No single install command. Read the steps in the README (opens in new tab)
- Maintenance
- Active
- Last commit 2 months ago
- Stars
- 438
- On GitHub
- Licence
- Apache-2.0
- Allows business use
- Downloads for
- None listed
- No release files
- Latest version
- None yet
- No releases
- Built with
- Python
- 96% of the code
Is it looked after?
- Contributors
- 1
- Last 90 days
- Issues
- None
- No issues filed
- Pull requests
- 0 merged
- 0 waiting
- Releases
- None
- Never published
- No contributing guide
- No security policy
- No code of conduct
Similar repos
| Repo | Stars | Last commit | Licence | Downloads for |
|---|---|---|---|---|
| adithya-s-k / FineEnvs | 554 | 12 days ago | Apache-2.0 | No downloads |
| oumi-ai / oumi | 9,429 | 3 days ago | Apache-2.0 | No downloads |
| opendilab / LightRFT | 379 | 2 months ago | Apache-2.0 | No downloads |
More: languages, topics and project record
Languages
- Python 96%
- Shell 4%
Topics
- grpo
- llm
- reinforcement-learning
- rlhf
Project record
- JobFine-tuning
- StatusListed
- Commits5
- Forks42
- Watching17
- Stars in 28 days0
- CreatedJan 7, 2026
- First seen by usNov 7, 2026
- CheckedNov 7, 2026