Update README.md

README.md

@@ -1,36 +1,14 @@
-# 🤝 GRPO Demo with Phi-2 — Helpfulness Reward
-
-This Space demonstrates **GRPO (Generative REINFORCE with Proximal Optimization)** —  
-the RL method used in DeepSeek-R1 — applied to **helpfulness alignment**.
-
-## 🚀 What happens here
-
-1. You enter a prompt  
-2. Phi-2 generates an answer  
-3. A **helpfulness reward function** scores it:
-   - detail  
-   - clear reasoning  
-   - helpful intent  
-   - no refusals  
-4. GRPO performs a policy gradient update:
-   \[
-   L = -R \log \pi + \beta KL(\pi || \pi_{\text{ref}})
-   \]
-5. The model improves step-by-step  
-6. You see a **reward curve** over time
-
-## 🧩 Files
-
-- `policy.py` — loads Phi-2  
-- `reward_fn.py` — helpfulness scoring  
-- `grpo_train.py` — GRPO update  
-- `app.py` — UI  
-- `requirements.txt`
-
+---
+title: Phi2 Helpfulness Grpo Demo
+emoji: 🐨
+colorFrom: pink
+colorTo: indigo
+sdk: gradio
+sdk_version: 6.0.2
+app_file: app.py
+pinned: false
+license: mit
+short_description: phi2-helpfulness-grpo-demo
 ---
 
-You can replace the reward with:
-- toxicity  
-- factuality  
-- math correctness  
-- chain-of-thought quality  
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference