AI Model Fine-tuning Framework
1.8k 2026-04-18
2U1/Qwen-VL-Series-Finetune
An open-source implementation for efficiently fine-tuning Alibaba Cloud's Qwen-VL series of multimodal large language models using HuggingFace and Liger-Kernel.
Core Features
Supports various fine-tuning methods (LoRA/QLoRA, Full-finetuning)
Enables multi-image and video training
Includes advanced optimization techniques (Deepspeed, Liger-Kernel)
Facilitates preference alignment (DPO, GRPO)
Handles mixed-modality datasets
Detailed Introduction
This project provides a comprehensive open-source framework for fine-tuning the Qwen-VL series of multimodal large language models developed by Alibaba Cloud. It leverages popular tools like HuggingFace and Liger-Kernel to offer efficient and flexible training capabilities. Researchers and developers can utilize this tool to adapt Qwen-VL models for specific tasks, supporting various fine-tuning strategies, multi-modal data inputs including video, and advanced optimization techniques, thereby extending the models' applicability across diverse AI applications.