有关训推差异和old-log-pro
- Dominant language
- Python
- Stars
- 3.4k
- Forks
- 312
- Avg merge
- 1h 2m
- Merged PRs (30d)
- 2
Description
如上。
您好,在我阅读调试了ROLL的代码后,我产生了如下问题:
1.训推差异:我注意到似乎ROLL并没有集成一些修复训推差异的方法如:TIS,MIS(mask掉的)
2.有关agentic pipline 中,无论是异步还是同步,似乎old-log-prob都是由当前的actor-trainer产生,并且结合我们通常的 ppo-epoch=1。这样似乎ratio总是维持着1?一些clip off-policy算法并没有起到作用?
我对ROLL框架的整理梳理可能不是很全,如有部分理解错误希望得到纠正。
Contributor guide
No contributing guide indexed for this repository
Research direction
No files, tests, or entry points are named. Start by tracing how old-log-prob is produced in the synchronous and asynchronous agentic pipelines, then examine PPO epoch handling and whether TIS or MIS is integrated. Done requires resolving whether the observed ratio behavior and training-inference discrepancy are actual bugs, with maintainer-confirmed scope and tests.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100