[Severity] High similarity with Colossal-AI about openrlhf HOT 4 CLOSED

We only referred to your directory file structure, and we developed everything else ourselves (Especially core RL technology and performance optimization). I don't think the directory structure and a few lines of comments have any real contribution.

from openrlhf.

Hello, our code is developed based on deepspeed and ray, while you used colossalai for development. Obviously it is completely impossible for us to be using the same technology stack.
Our framework uses step-wise RL, whereas yours uses single-step RL.
The core of a framework is performance optimization and hyperparameter tuning & PPO implementation details, not the directory structures/comments.
Finally, your colossalchat code also includes contributions from us, including PPO tuning and ray.

from openrlhf.

Colossalchat Team contacted me earlier about RL issues

The wechat history proves that colossalchat's core RL technology was supported by me.

Our developers also contributed the ray components to colossalaichat
https://github.com/hpcaitech/ColossalAI/pull/3309/files

from openrlhf.

@binmakeswell
Thank you for the discussion.
Let's keep in touch in the future and work together to improve the RLHF ecosystem.

from openrlhf.