Add metadata and improve model card

Hi! I'm Niels from the community science team at Hugging Face.

I've opened this PR to improve the model card for DRA-GRPO. Specifically, I have:
- Added metadata for `pipeline_tag`, `library_name`, and `base_model` to make the model more discoverable.
- Improved the markdown structure for better readability.
- Linked the model to its corresponding paper and code repository.
- Added a BibTeX citation.

These changes will enable features like the "Use in Transformers" button and ensure the model appears in the correct categories on the Hub.

README.md

@@ -1,7 +1,38 @@
 ---
 license: cc-by-4.0
+library_name: transformers
+pipeline_tag: text-generation
+base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
+tags:
+- mathematics
+- grpo
+- reinforcement-learning
+- reasoning
 ---
 
-This model is described in the paper [DRA-GRPO: Exploring Diversity-Aware Reward Adjustment for R1-Zero-Like Training of Large Language Models](https://arxiv.org/abs/2505.09655).
+# DRA-GRPO
 
-Full code is in: https://github.com/xiwenc1/DRA-GRPO
+This repository contains the weights for DRA-GRPO, a model developed to enhance mathematical reasoning in Large Language Models (LLMs) through **Diversity-aware Reward Adjustment (DRA)**. 
+
+The model is based on the [DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) architecture and was trained using a framework that calibrates reward signals using semantic density to prevent policy collapse into redundant reasoning paths.
+
+## Resources
+- **Paper:** [DRA-GRPO: Your GRPO Needs to Know Diverse Reasoning Paths for Mathematical Reasoning](https://huggingface.co/papers/2505.09655)
+- **ArXiv:** [2505.09655](https://arxiv.org/abs/2505.09655)
+- **Code:** [xiwenc1/DRA-GRPO](https://github.com/xiwenc1/DRA-GRPO)
+
+## Method Overview
+DRA-GRPO implements Diversity-aware Reward Adjustment (DRA), a theoretically grounded framework that uses Submodular Mutual Information (SMI) and Inverse Propensity Scoring (IPS). This mechanism de-biases gradient estimation by creating a repulsive force against redundancy, driving the policy to achieve better coverage of the high-reward reasoning landscape.
+
+## Performance
+Empirical evaluations demonstrate that DRA-GRPO consistently outperforms standard GRPO baselines, achieving an average accuracy of 58.2% on mathematical benchmarks (such as MATH-500, AMC23, and AIME24) with high data efficiency.
+
+## Citation
+```bibtex
+@article{chen2025dra,
+  title={DRA-GRPO: Your GRPO Needs to Know Diverse Reasoning Paths for Mathematical Reasoning},
+  author={Xiwen Chen and Wenhui Zhu and Peijie Qiu and Xuanzhao Dong and Hao Wang and Haiyu Wu and Huayu Li and Aristeidis Sotiras and Yalin Wang and Abolfazl Razi},
+  journal={arXiv preprint arXiv:2505.09655},
+  year={2025}
+}
+```