Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
轻量级 LLM Post-training 框架，支持 SFT、RLVR、On-Policy KD、Guide KD 及混合训练；实现单轮/多轮 Guide 蒸馏、多教师蒸馏、Reward 混合训练与自动化数据分流👩🎓👨🎓
| Date | Stars |
|---|---|
| 2026-07-31 | 938 |
| 2026-08-06 | 939 |
Today
+1 stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# 🥋 LLM-Dojo > A lightweight playground for `RLHF` and `SFT` experiments, with support for `RLVR`, `KD`, and `Guide-KD`. > > 轻量级 RLHF/SFT 实验平台，支持 `RLVR`、`KD` 与 `Guide-KD`。 ## 📋 Overview | 模块 | 说明 | |------|------| | [`openrlhf-kd`](./openrlhf-kd/) | 当前主线，基于 OpenRLHF 重构，实现 `RLVR` + `KD` + `Guide-KD` | | [`main_train.py`](./main_train.py) | 简洁 `SFT` 训练入口 | ## 🎯 RLVR [`openrlhf-kd`](./openrlhf-kd/) 是这个仓库当前最核心的部分，基于 OpenRLHF 构建，具体训练使用可参见文档 [`openrlhf-kd/examples/README.md`](./openrlhf-kd/examples/README.md) **主要改动：** 1. 精简框架，只保留 `RLVR` 部分，移除了 `critic` 等不需要的内容 2. 增加 `KD`、`Guide-KD` 与 `reward` 的混合训练，支持按 `datasource` 路由 ## ✏️ SFT 根目录的 `SFT` 部分保持了比较简洁的训练入口，适合快速微调实验。 **特性：** - 支持 `Deepspeed` - 支持 `LoRA`、`QLoRA`、全参微调 - 自动适配 chat template 示例文件可参见 `data/sft_data.jsonl` **Quick Start：** ```bash bash run_example.sh ``` 或： ```bash deepspeed --include localhost:0,1 main_train.py \ --train_data_path /path/to/data.jsonl \ --model_name_or_path /path/to/model \ --task_type sft \ --train_mode qlora \ --output_dir /path/to/output ```
Excerpt of 1,061 characters
Read on GitHub
264
Nankai University · China
4
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:aed510557ca52e93, llm:description: '轻量级 LLM Post-training 框架，支持 SFT、RLVR、On-Policy KD、Guide KD 及混合训练；实现单轮/多轮 Guide 蒸馏、多教师蒸馏、Reward 混合训练与自动化数据分流'
matched fp:aed510557ca52e93, llm:description: '轻量级 LLM Post-training 框架，支持 SFT、RLVR、On-Policy KD、Guide KD 及混合训练；实现单轮/多轮 Guide 蒸馏、多教师蒸馏、Reward 混合训练与自动化数据分流'