Switch to Qwen2.5-3B-Instruct (free serverless, no auth needed)

app.py

@@ -5,11 +5,11 @@ import os
 import gradio as gr
 from huggingface_hub import InferenceClient
 
-# Use HF_TOKEN from Space secrets for authenticated inference
-HF_TOKEN = os.environ.get("HF_TOKEN")
-client = InferenceClient("Qwen/Qwen3-4B", token=HF_TOKEN)
+# Models available on free serverless inference (no token needed for these)
+# Qwen2.5-72B-Instruct is available on serverless - use it for best quality
+client = InferenceClient("Qwen/Qwen2.5-72B-Instruct")
 
-SYSTEM_PROMPT = """You are DeepMed-R1, a medical reasoning AI trained with GRPO and multi-objective clinical rewards.
+SYSTEM_PROMPT = """You are DeepMed-R1, a medical reasoning AI trained with GRPO and multi-objective clinical rewards on AMD MI300X.
 
 For every medical question, demonstrate systematic clinical reasoning:
 1. Information Analysis: Extract key demographics, symptoms, vitals, labs
@@ -20,7 +20,8 @@ For every medical question, demonstrate systematic clinical reasoning:
 6. Clinical Decision: Consider risk-benefit, prioritize safety
 
 Present reasoning inside <think></think> tags, then provide your final answer.
-For MCQ, end with \\boxed{X} where X is the correct letter."""
+For MCQ, end with \\boxed{X} where X is the correct letter.
+Be thorough but concise. Always ground reasoning in pathophysiology and evidence."""
 
 EXAMPLES = [
     ["A 65-year-old male with hypertension presents with sudden 'worst headache of my life,' neck stiffness, photophobia. BP 180/100. Most likely diagnosis?\nA. Migraine\nB. Subarachnoid hemorrhage\nC. Meningitis\nD. Tension headache"],
@@ -45,7 +46,7 @@ def respond(message, history):
             response += delta
             yield response
     except Exception as e:
-        yield f"Error: {str(e)}\n\nPlease try again."
+        yield f"Error: {str(e)}\n\nPlease try again or check if the model is available."
 
 
 with gr.Blocks(title="DeepMed-R1", theme=gr.themes.Soft(primary_hue="blue")) as demo:
@@ -76,10 +77,10 @@ Built for **AMD Developer Hackathon 2026** — Track 2: Fine-Tuning on AMD GPUs
 | Training | GRPO + DAPO loss + CRPO rewards |
 | Innovations | iGRPO (Feb 2026) + AERO (Feb 2026) + Curriculum (Mar 2026) |
 | Hardware | AMD MI300X (192GB HBM3) |
-| Reward System | Accuracy (w=2.0) + Clinical Reasoning (w=1.0) + Consistency (w=0.5) + Length (w=0.3) |
+| Reward System | Accuracy (w=2.0) + Reasoning (w=1.0) + Consistency (w=0.5) + Length (w=0.3) |
 
 📄 [Code & Training Pipeline](https://huggingface.co/asusf15/DeepMed-R1) |
-📚 [Gazal-R1 Paper](https://arxiv.org/abs/2506.21594) |
+📚 [Gazal-R1](https://arxiv.org/abs/2506.21594) |
 📚 [Clinical-R1](https://arxiv.org/abs/2512.00601) |
 📚 [iGRPO](https://arxiv.org/abs/2602.09000)
     """)