agentscope-ai / agentscope-ai/Trinity-RFT

[roadmap] Trinity-RFT development

未关闭
#51 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看
主要语言
Python
星标
701
派生
79
平均合并
8 小时 7 分钟
30 天内合并 PR
1

描述

- [x] Support generic multi-step RL scenarios (infra & algorithm)
- [ ] Further optimize utilization of computational resources
- [ ] Support automatic load balancing
- [x] Improve robustness of `WorkflowRunner` --> Upgraded to `Scheduler`
- [x] Support training with Megatron-LM
- [x] Implement more algorithms for off-policy / asynchronous RL
- [x] Implement more advanced sampling strategies for task / experience buffer
- [x] Further integrate RL process with advanced data processing functionalities
- [ ] Continue refining Trinity-Studio

贡献指南

打开贡献指南

评估

这个 Issue 还没有评估数据。

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。