Feature: Dr. GRPO Support
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 10.8k
- Forks
- 989
- Avg merge
- 6h 29m
- Merged PRs (30d)
- 85
Description
It would be nice if art would support the Dr. GRPO optimizer.
Basically, Dr. GRPO is supposed to remove the response length bias which is present in the original GRPO, while retaining the performance benefits of GRPO. See Dr. GRPO Paper for more info.
Important Note: trl's GRPOTrainer already supports this by setting loss_type="dr_grpo" in GRPOConfig.
I believe this technique might be beneficial, especially for multi-turn RL scenarios, in which token length increase is more significant.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
No ART files or tests are named. Start by locating ART's GRPO training entry point and comparing its loss configuration with trl's GRPOTrainer support for loss_type="dr_grpo" in GRPOConfig, then read the linked Dr. GRPO paper. Done means ART can select Dr. GRPO and its behavior is covered by appropriate tests.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 35/100