joykirat

/

Qwen3-4B-TRAAC

like 1

This PR significantly enhances the model card for the TRAAC (Think Right with Adaptive, Attentive Compression) Qwen3-4B model by populating it with detailed information from the paper Think Right: Learning to Mitigate Under-Over Thinking via Adaptive, Attentive Compression and its corresponding GitHub repository (https://github.com/joykirat18/TRAAC).

Key improvements include:

• Updating metadata with pipeline_tag: text-generation, license: mit, and relevant tags (reasoning, qwen3, math) for better discoverability and functionality on the Hub.
• Providing a comprehensive model description, outlining its purpose, mechanism (adaptive, attentive compression), and key performance highlights.
• Including a visual overview of the TRAAC framework with the architectural diagram from the GitHub repository.
• Populating model details such as developers, model type, language, and base model.
• Detailing intended uses, out-of-scope uses, biases, risks, and limitations.
• Providing information about training and evaluation details, directing users to the GitHub repository for scripts and full instructions.
• Adding the official BibTeX citation.

Please note that a direct, copy-pastable Python code snippet for basic inference using the transformers library has not been included in the "How to Get Started" section. This is because the provided GitHub README explicitly marks this section as [More Information Needed] and primarily points to a complex setup involving the verl library and specific scripts for running the model, which goes beyond a simple, ready-to-use transformers inference example directly found in the source. Users are directed to the GitHub repository for full details on installation and usage.