MMR1

/

MMR1-7B-SFT

like 0

Follow

MMR1 12

This PR significantly improves the model card for the MMR1 model, presented in MMR1: Enhancing Multimodal Reasoning with Variance-Aware Sampling and Open Resources.

Key improvements include:

• Metadata: Addition of license: apache-2.0, pipeline_tag: image-text-to-text, library_name: transformers, and relevant tags (qwen2_5_vl, multimodal-llm, multimodal-reasoning, math-reasoning).
• Contextual Metadata: Inclusion of datasets (MMR1/MMR1-SFT, MMR1/MMR1-RL) and base_model (Qwen/Qwen2.5-VL-7B-Instruct) for better lineage and discoverability.
• Performance Metrics: Addition of model-index with an average performance score for the MMR1-7B-RL variant on mathematics-related multimodal reasoning benchmarks.
• Content Enrichment: Integration of the paper abstract, a direct link to the GitHub repository, and key sections from the GitHub README (Introduction, Methodology, Open Resources, Evaluation, Analysis, Qualitative Demo, Acknowledgement, Citation, and License) for a more informative model overview. Relevant images from the GitHub repository are also included.
• Usage Example: Addition of a clear "Quick Start (Inference)" code snippet demonstrating how to use the model with the transformers library, which directly supports the library_name metadata and enables the automated "How to use" widget. The example is designed for image-text input, aligning with the model's capabilities.
• Clean-up: Removal of internal "File information" from the model card.

This enhances the model's discoverability, usability, and documentation for researchers and users on the Hugging Face Hub.