NeMo RL + Gym integration for SFT
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 1.2k
- Forks
- 349
- Avg merge
- 1d 21h
- Merged PRs (30d)
- 318
Description
Use cases, pain points, and background
Gym can be used for offline rollout collection for SFT/DPO. I don't think we need to change anything on the NeMo RL infra side, I think we might just need some utils that convert from Responses to NeMo RL message logs. But we need to pipeclean the flow and have a tutorial in place.
Description:
What should we do?
Design:
What files should be touched? What logic should be written?
Out of scope:
What are some items that this issue could be mistaken to cover that this issue should explicitly NOT cover?
Acceptance Criteria:
- Collect rollouts using Gym
- Perform SFT on them and show that performance improves on the appropriate validation set
- Add your W&B in the PR.
- Contribute a conversion script, if any.
- Write a tutorial for how to do this with NeMo RL
Thanks to @HeyyyyyyG for the raise!
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 files, tests, or entry points are named. Start by locating the Gym rollout-collection flow and the NeMo RL message-log and SFT interfaces. Done means rollouts are collected, SFT improves the appropriate validation result, any conversion script and W&B reporting are included, and a NeMo RL tutorial documents the workflow.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- documentation, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 20/100