LMMs-Lab-Turtle

/

Qwen-2.5VL-7B-Cold-Start

like 0

Follow

Turtle-Lab 5

This PR significantly improves the model card for the Osilly/Vision-R1-7B model, which corresponds to the Vision-SR1-7B model introduced in the paper "Self-Rewarding Vision-Language Model via Reasoning Decomposition".

The changes include:

• Adding library_name: transformers to the metadata. This enables the automated "Use in Transformers" code snippet, leveraging the detected Qwen2_5_VLForConditionalGeneration architecture and transformers dependency mentioned in the config.json and GitHub README.
• Adding pipeline_tag: image-text-to-text to the metadata, which ensures the model is discoverable under the appropriate task on the Hugging Face Hub (e.g., https://huggingface.co/models?pipeline_tag=image-text-to-text). This aligns with the model's functionality as a Vision-Language Model.
• Integrating the paper's abstract and key information from the GitHub README to provide a comprehensive "About Vision-SR1" section.
• Including direct links to the official paper (https://huggingface.co/papers/2508.19652) and the GitHub repository (https://github.com/zli12321/Vision-SR1).
• Embedding relevant figures from the GitHub repository to visually explain the method and dataset.
• Adding links to associated models and datasets on the Hugging Face Hub, as found in the GitHub README.
• Providing installation and training setup instructions based on the GitHub repository.
• Noting the incorrect BibTeX citation in the original GitHub README for the Vision-SR1 paper and including the correct citation for the EasyR1 codebase.

Please review these updates.