om-ai-lab/VLM-R1

Implements GRPO (Group Relative Policy Optimization) reinforcement learning on vision-language models like Qwen2.5-VL and InternVL, demonstrating superior out-of-domain generalization compared to SFT approaches on tasks like referring expression comprehension and open-vocabulary detection. Supports flexible training configurations including full fine-tuning, LoRA, multi-node distributed training, and multi-image inputs, with customizable reward functions for different vision tasks. Provides optimized inference implementations via xllm and vllm-ascend frameworks for hardware deployment on Ascend accelerators.

5,864 stars. Actively maintained with 1 commit in the last 30 days.