alibaba / alibaba/ROLL

Does ROLL support DAPO training for LLMs or VLMs?

Open
#103 1 comment 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
3.4k
Forks
312
Avg merge
1h 2m
Merged PRs (30d)
2

Description

I noticed that in `ROLL/roll/pipeline/rlvr/rewards/math_rule_reward_worker.py`, there are some parameters related to DAPO, such as `epsilon_high`, `mask_truncated_completions`, and `loss_type`.

Image

Image

Image

However, I couldn’t find any description of these parameters in the official documentation or in `ROLL/docs_roll/docs/#U7b80#U4f53#U4e2d#U6587/#U5feb#U901f#U5f00#U59cb/config_guide_cn.md`.

Does ROLL currently support DAPO? If so, how should I use these parameters?

Contributor guide

No contributing guide indexed for this repository

Research direction

Start with ROLL/roll/pipeline/rlvr/rewards/math_rule_reward_worker.py and inspect how epsilon_high, mask_truncated_completions, and loss_type are used. Compare those parameters with docs_roll/docs/#U7b80#U4f53#U4e2d#U6587/#U5feb#U901f#U5f00#U59cb/config_guide_cn.md. Document whether DAPO is supported and explain how these parameters should be configured, if their behavior is established.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
documentation
Issue type
Documentation
Difficulty
3/5
Estimated time
1-2 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
30/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.