grpo

Use PEFT or Full-parameter to CPT/SFT/DPO/GRPO 500+ LLMs (Qwen2.5, Llama4, InternLM3, GLM4, Mistral, Yi1.5, DeepSeek-R1, ...) and 200+ MLLMs (Qwen2.5-VL, Qwen2.5-Omni, Qwen2-Audio, Ovis2, InternVL3, Llava, MiniCPM-V-2.6, GLM4v, Xcomposer2.5, DeepSeek-VL2, Phi4, GOT-OCR2, ...).

Solve Visual Understanding with Reinforced VLMs

Explore the Multimodal “Aha Moment” on 2B Model

Collect every awesome work about r1!

🌾 OAT: A research-friendly framework for LLM online alignment, including preference learning, reinforcement learning, etc.

The open source implementation of DeepSeek-R1. 开源复现 DeepSeek-R1

Unleashing the Power of VLMs in Autonomous Driving via Reinforcement Learning and Reasoning

OpenVLThinker: An Early Exploration to Vision-Language Reasoning via Iterative Self-Improvement

🐭 A tiny single-file implementation of Group Relative Policy Optimization (GRPO) as introduced by the DeepSeekMath paper

Aims for memory-efficient training (24GB VRAM) on consumer GPUs. Optimizing language models through guidance tokens in reasoning chains, based on DeepSeekRL-Extended.

Recreating the minimal training methods of DeepSeek-R1 for small langauge models.

An open source implementation of R1

Simple repository for training small reasoning models

Distributed Reinforcement Learning for LLM Fine-Tuning with multi-GPU utilization

A travel agent based on Qwen2.5, fine-tuned by SFT + DPO/PPO/GRPO using traveling question-answer dataset, a mindmap can be output using the response. A RAG system is build upon the tuned qwen2, using Prompt-Template + Tool-Use + Chroma embedding database + LangChain

A reinforcement learning agent that learns to solve mazes using Group Relative Policy Optimization (GRPO).

使用trl、peft、transformers等库，实现对huggingface上模型的微调。

This repository contains blog post for GRPO RL algorithm using simple Grid World environment.

LLM finetuning for Sudoku solving

Reinforcement Fine-Tuning