dongboklee

/

dPRM-14B

like 0

This PR significantly enhances the model card for the adapter based on deepseek-ai/DeepSeek-R1-Distill-Qwen-14B, which is part of the work presented in the paper "Rethinking Reward Models for Multi-Domain Test-Time Scaling".

Key improvements include:

• Adding the license: apache-2.0 tag to the metadata, specifying the license for the model code (datasets use CC-BY-NC-4.0).
• Adding a reward-model tag for better discoverability.
• Populating the "Model Details" section with a comprehensive description derived from the paper's abstract.
• Including direct links to the paper and the GitHub repository.
• Adding a "How to Get Started" section with environment setup and an inference code snippet for generative reward models, directly sourced from the GitHub README.
• Providing information on uses, limitations, training, and evaluation, linking to the GitHub repository where appropriate.
• Including the BibTeX citation.

The library_name remains peft as the model is explicitly a PEFT (LoRA) adapter, and pipeline_tag remains text-generation given the model's capability for generative verification rationales and its CAUSAL_LM task type.

Please review and merge this PR.