aimagelab

/

LLaVA_MORE-gemma_2_9b-finetuning

like 0

Follow

AImageLab 26

This PR significantly enhances the model card for aimagelab/LLaVA_MORE-gemma_2_9b-finetuning by providing comprehensive details and improving discoverability.

Key updates include:

• Metadata: Added pipeline_tag: image-text-to-text, a comprehensive set of tags, base_model (google/gemma-2-9b-it), and datasets (liuhaotian/LLaVA-Pretrain, liuhaotian/LLaVA-Instruct-150K) for better categorization and searchability. The license is explicitly set to Apache 2.0.
• Model Details & Description: Replaced placeholder content with a detailed overview of the LLaVA-MORE family, including its purpose, LLM and visual backbone variations explored, and specific details for this model variant (Gemma-2 9B + CLIP).
• Model Sources: Added direct links to the paper on Hugging Face Papers, the official GitHub repository, the project page, the Hugging Face collection, and a general Hugging Face Space demo.
• Usage Example: Included a ready-to-use Python code snippet for inference, specifically tailored for this model variant, with guidance on handling out-of-memory issues.
• Performance Benchmarks: Integrated the detailed performance table and plot from the original GitHub repository, showcasing the model's evaluation results.
• Training Details: Provided information on the two-stage training process, including datasets and procedure.
• Checkpoints: Included the full table of all LLaVA-MORE checkpoints with their respective Hugging Face links.
• Latest Updates: Included the "Latest Updates" section from the GitHub repository to keep users informed about project milestones.
• Citation & Acknowledgments: Ensured the BibTeX citation is present and retained the acknowledgments section.

This update aims to make the model card a much richer, more informative, and user-friendly resource on the Hugging Face Hub.