EvolvingLMMs-Lab/open-r1-multimodal

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

1603 stars · 74 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

A fork of Hugging Face's open-r1 that extends the R1 GRPO reinforcement learning training paradigm to multimodal (vision-language) models like Qwen2-VL and Aria-MoE. It includes training scripts, a curated 8k verifiable math reasoning dataset, and released GRPO-trained model checkpoints.

Use cases

• train a vision-language model with GRPO reinforcement learning
• reproduce R1-style reasoning training on multimodal models
• fine-tune Qwen2-VL with verifiable-reward RL
• create custom multimodal RL training data from math datasets
• study reward dynamics of multimodal reasoning models

When to choose

• you want to apply GRPO/R1-style RL training to vision-language models
• you need a starting point for multimodal reasoning model research
• you want the released 8k verified multimodal math RL dataset

When to avoid

• you need efficient large-scale training - the framework is slow (10 hours per epoch on 8 H100s for 8k samples)
• you only need text-only LLM training - upstream open-r1 covers that
• you need a production-ready, stable training pipeline

Facets

library · maturity experimental

llm-training machine-learning deep-learning large-language-models machine-learning deep-learning computer-vision python grpo reinforcement-learning vision-language-models multimodal reasoning qwen2-vl rlhf training-scripts gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="EvolvingLMMs-Lab/open-r1-multimodal")