Thunderbird2410/grpo-vl-lora

Model Card: GRPO-Optimized Qwen3-VL (LoRA)

This repository hosts the LoRA adapters for a version of Qwen/Qwen3-VL-8B-Instruct and fine-tuned via Group Relative Policy Optimization (GRPO).

Training procedure

This model was trained with GRPO, a method introduced in DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.

Reward Functions

The policy was guided by three specific reward signals:

1. Format Compliance: Enforcing strict usage of <reasoning> and <answer> XML tags.
2. Sequence Logic: Validating that the reasoning block precedes the answer block.
3. Mathematical Correctness: Exact-match validation of vector/matrix outputs against ground truth (MathVista dataset).

Environment

• Hardware: AMD Instinct MI300X rocm 6.4
• Framework: PyTorch + TRL + vLLM
• Base Model: Qwen/Qwen3-VL-8B-Instruct

Framework versions

• TRL: 0.26.2
• Transformers: 4.57.3
• Pytorch: 2.9.1+git8907517
• Datasets: 4.4.2
• Tokenizers: 0.22.1

Citations

Cite GRPO as:

@article{shao2024deepseekmath,
    title        = {{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
    author       = {Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo},
    year         = 2024,
    eprint       = {arXiv:2402.03300},
}

Cite TRL as:

@misc{vonwerra2022trl,
    title        = {{TRL: Transformer Reinforcement Learning}},
    author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
    year         = 2020,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/huggingface/trl}}
}