Zero-Qwen-VL
AI developer platformsZero-Qwen-VL is an open-source project aimed at building from scratch a LLaVA-architecture image-text model with better support for Chinese. The project uses a combination of Qwen-VL's vision encoder and the Qwen2-0.5B-Instruct language model, and has open-sourced the training code, data, and three-stage model weights. Its training data includes LLaVA-CC3M-Pretrain-595K and LLaVA-Instruct-150K from Chinese-LLaVA-Vision-Instructions, as well as translated TextOCR-GPT4o and ShareGPT-4o data. The project provides complete environment configuration, weight extraction and integration, training, and testing procedures, making it suitable for developers and researchers who wish to study multimodal model composition and Chinese image-text training.