hiyouga/EasyR1

Health v2 · maintenance only

Adoption not part of the score

5129 stars · 387 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

EasyR1 is an efficient, scalable reinforcement learning training framework for large language models and vision-language models, built as a fork of veRL. It supports algorithms like GRPO, DAPO, and GSPO, leveraging HybridEngine and vLLM's SPMD mode for performance.

Use cases

• train an LLM with GRPO reinforcement learning
• fine-tune Qwen vision-language models with RL
• run RLHF post-training on DeepSeek-R1 distill models
• scale multi-modal RL training across GPUs
• apply LoRA-based reinforcement learning to a 7B model
• train language models with DAPO or GSPO algorithms

When to choose

• you need RL post-training (GRPO/DAPO/GSPO) for LLMs or VLMs
• you want a veRL-based framework with vision-language model support
• you need scalable multi-GPU RL training with vLLM rollouts
• you want checkpointing, experiment tracking, and LoRA in one RL framework

When to avoid

• you only need supervised fine-tuning without RL
• you lack multi-GPU hardware or Docker support
• you need a simple inference or serving solution rather than training
• you prefer a framework with broader model coverage than the supported Qwen/Llama/DeepSeek families

Facets

framework · maturity active

llm-training reinforcement-learning machine-learning gpu-computing large-language-models reinforcement-learning machine-learning deep-learning python grpo dapo verl vllm vision-language-models rlhf qwen deepseek lora multi-modal gpu docker linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="hiyouga/EasyR1")