2U1/Qwen-VL-Series-Finetune

An open-source implementaion for fine-tuning Qwen-VL series by Alibaba Cloud.

What it solves

This project provides a comprehensive toolkit for fine-tuning the Qwen-VL series of vision-language models (including Qwen2-VL, Qwen2.5-VL, Qwen3-VL, and Qwen3.5). It simplifies the process of adapting these multimodal models to specific tasks, supporting various training methods from full fine-tuning to parameter-efficient techniques like LoRA and QLoRA.

How it works

The repository provides training scripts built on HuggingFace and Liger-Kernel to optimize memory usage and speed. It supports multiple training paradigms:

  • Supervised Fine-Tuning (SFT): Standard training on labeled datasets.
  • Preference Optimization: Implements Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO) to align model behavior.
  • Reasoning Training: Supports explicit reasoning traces (thinking process) for Qwen3-VL-Thinking and Qwen3.5 models using a specialized dataset format.
  • Classification: Includes scripts for fine-tuning the models for classification tasks.

It handles diverse data modalities, including single images, multi-image sets, and videos, following the LLaVA dataset specification.

Who it’s for

  • AI researchers and developers looking to customize Qwen-VL models for domain-specific visual understanding.
  • Users wanting to implement advanced alignment techniques like GRPO or DPO on multimodal models.
  • Developers needing to train vision-language models with limited hardware resources via LoRA/QLoRA and Liger-Kernel optimization.

Highlights

  • Broad Model Support: Compatible with Qwen2-VL, Qwen2.5-VL, Qwen3-VL (including MoE variants), and Qwen3.5.
  • Optimization: Integrates Liger-Kernel for memory efficiency and fused MoE expert kernels for MoE models.
  • Flexible Training: Supports full fine-tuning, LoRA, QLoRA, and the ability to selectively unfreeze specific layers of the vision tower or LLM.
  • Multimodal Capabilities: Native support for multi-image and video training.
  • Advanced Alignment: Built-in support for DPO and GRPO training methods.

Related

  • Project
    om-ai-lab/VLM-R1

    VLM-R1 is a framework for training R1-style Large Vision-Language Models using GRPO to improve reasoning and generalization in tasks like visual grounding, math, and object detection.

  • Dispatch
    Vision Language Model Alignment in TRL

    Hugging Face has expanded the TRL library to support advanced alignment methods for Vision Language Models, including MPO, GRPO, and GSPO, alongside native SFT support and vLLM integration.

  • Project
    hiyouga/EasyR1

    An efficient, scalable reinforcement learning training framework that supports multi-modality models and various RL algorithms like GRPO.

  • Dispatch
    Qwen2.5-VL-32B Release Notes

    Qwen has released Qwen2.5-VL-32B-Instruct, a vision-language model optimized via reinforcement learning for superior mathematical reasoning, fine-grained image understanding, and human-aligned responses.

  • Dispatch
    Hugging Face Vision Language Models 2025 Update

    Hugging Face provides a comprehensive overview of the 2024-2025 evolution of Vision Language Models (VLMs), highlighting trends in any-to-any architectures, reasoning models, and the rise of Vision-Language-Action (VLA) models for robotics.