#4646th of 53 in Training and fine-tuning
vlm-r1
GRPO reinforcement-learning training for Qwen2.5-VL and InternVL models
- Stars
- 6k
- License
- Apache-2.0
- Last commit
- Jul 2026
- Last release
- Apr 2025
- Language
- Python
5,889 on 20 Apr 20266,045 on 9 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
VLM-R1 applies R1-style GRPO reinforcement learning to vision-language models, with scripts for full fine-tuning, LoRA, multi-node and multi-image training. It supports QwenVL and InternVL, loads custom jsonl data, and ships 3B checkpoints trained for referring expression comprehension, open-vocabulary detection and multimodal math. An SFT baseline via LLaMA-Factory and Ascend inference guides are included.
Who it is for: ML engineers fine-tuning vision-language models with GRPO on their own GPUs
Strengths
- Full, LoRA, frozen-vision, multi-node and multi-image GRPO training with ready run scripts
- Released 3B checkpoints for OVD, Math and REC, plus Hugging Face demos
- Custom jsonl loader and documented steps for adding new VLMs and reward functions
- Includes SFT comparison setup and Ascend 910B/300I Duo inference guides
Weaknesses
- Recipes centre on REC, OVD and a few tasks; other tasks need your own reward functions
- Last tagged release is 2025-04-15; later changes exist only as commits
- Setup is a conda Python 3.10 env plus setup.sh; no compose file
- Only 3B checkpoints are listed; larger model results are not documented
What it needs
- GPU required
- Docker
- Needs PyTorch, DeepSpeed, LLaMA-Factory (SFT baseline), Hugging Face datasets
- Models: Qwen2.5-VL, InternVL
What changed
From the release notes. No flag means the note does not say so, not that the upgrade is safe.
Quick answers
Is vlm-r1 open source?
Yes. Its code is on GitHub under the Apache-2.0 license.
Can you run vlm-r1 with Docker?
Yes. Its repository has a Docker setup: a Dockerfile, a Compose file or a published image.
Does vlm-r1 need a GPU?
Yes, it needs a GPU.
Also in Training and fine-tuning
See all 53| Rank | Project | Score |
|---|---|---|
| 1 | unslothDesktop and web app to run and fine-tune local models | 88 out of 100 |
| 2 | peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models | 88 out of 100 |
| 3 | deepspeedPyTorch library for distributed training of very large models with ZeRO | 86 out of 100 |
| #4 | verlReinforcement learning training library for LLM post-training with PPO and GRPO | 86 out of 100 |
| #5 | trlPost-training library with SFT, DPO, GRPO and KTO trainers | 86 out of 100 |