rwightman

/

timm-optim-caution

like 9

This PR improves the model card for the MARS optimizer by:

• Adding pipeline_tag: image-classification to reflect the optimizer's evaluation on relevant vision tasks like mini-imagenet and CIFAR, helping users discover it when filtering for this pipeline.
• Adding library_name: transformers, as the optimizer is used to train models from the Hugging Face Transformers library (e.g., GPT-2), enabling a predefined code snippet for usage.
• Updating the main title of the model card to the paper's official title and linking it to the Hugging Face paper page for clearer attribution.
• Adding a prominent link to the GitHub repository for easy access to the source code.
• Expanding the model card content by integrating comprehensive sections from the official GitHub README, including "About MARS," "Instantiations," detailed "Performance Comparisons" across various tasks (LLMs and Vision), "Training GPT-2 from Scratch," and instructions for "Reproducing Results."
• Including a "Customized Training" section with a Python code snippet from the official repository, demonstrating how to integrate the MARS optimizer into a PyTorch training loop.
• Adding the official Citation and Acknowledgements.

These updates provide a more complete, discoverable, and user-friendly overview of the MARS optimizer for the Hugging Face community.