Fix: remove assistant prefill, use simple robust approach

app.py

@@ -10,39 +10,29 @@ client = InferenceClient(token=HF_TOKEN)
 
 MODEL = "Qwen/Qwen2.5-72B-Instruct"
 
-SYSTEM_PROMPT = """You are DeepMed-R1, a medical reasoning AI trained with GRPO on AMD MI300X. 
+SYSTEM_PROMPT = """You are DeepMed-R1, a medical reasoning AI trained with GRPO on AMD MI300X.
 
-You ALWAYS respond in this EXACT format — no exceptions:
+For EVERY question you MUST follow this format exactly:
 
-<think>
-## 1. Information Analysis
-[Extract ALL key clinical features: age, sex, history, symptoms, vitals, labs, imaging]
+**Step 1 - Key Findings:**
+[List the important clinical features from the question]
 
-## 2. Differential Diagnosis  
-[List 3-5 possible diagnoses ranked by probability with brief justification]
+**Step 2 - Differential Diagnosis:**
+[List 3-4 possible diagnoses ranked by likelihood]
 
-## 3. Pathophysiology
-[Explain the underlying disease mechanism for the most likely diagnosis]
+**Step 3 - Pathophysiology:**
+[Explain WHY the most likely diagnosis causes these symptoms]
 
-## 4. Evidence-Based Reasoning
-[Reference clinical criteria, guidelines, or classic presentations]
+**Step 4 - Option Analysis:**
+- A: [evaluate this option]
+- B: [evaluate this option]
+- C: [evaluate this option]
+- D: [evaluate this option]
 
-## 5. Option-by-Option Elimination
-- A. [option text] → [why correct or incorrect]
-- B. [option text] → [why correct or incorrect]  
-- C. [option text] → [why correct or incorrect]
-- D. [option text] → [why correct or incorrect]
-</think>
+**Step 5 - Final Answer:**
+The answer is **X** because [one sentence justification].
 
-**Assessment:** [2-3 sentence clinical conclusion]
-
-**Final Answer:** \\boxed{X}
-
-Rules:
-- The <think> section MUST appear FIRST and contain ALL 5 numbered sections
-- Each section must have at least 2-3 sentences of clinical detail
-- Be specific: name diseases, mechanisms, criteria by proper medical terms
-- The reasoning must JUSTIFY the final answer logically"""
+IMPORTANT: Always show ALL 5 steps. Never skip steps. Be detailed and specific."""
 
 
 EXAMPLES = [
@@ -59,12 +49,9 @@ def respond(message, history):
     for h in history:
         if h[0]: messages.append({"role": "user", "content": h[0]})
         if h[1]: messages.append({"role": "assistant", "content": h[1]})
-    
     messages.append({"role": "user", "content": message})
-    # Prefill the assistant response to force it to start with <think>
-    messages.append({"role": "assistant", "content": "<think>\n## 1. Information Analysis\n"})
 
-    response = "<think>\n## 1. Information Analysis\n"
+    response = ""
     try:
         stream = client.chat_completion(
             model=MODEL,
@@ -80,11 +67,10 @@ def respond(message, history):
                 if hasattr(delta, "content") and delta.content:
                     response += delta.content
                     yield response
+        if not response:
+            yield "No response received. Please try again."
     except Exception as e:
-        if not response or response == "<think>\n## 1. Information Analysis\n":
-            yield f"⚠️ Error: {str(e)}"
-        else:
-            yield response
+        yield f"⚠️ Error: {str(e)}"
 
 
 with gr.Blocks(title="DeepMed-R1", theme=gr.themes.Soft(primary_hue="blue")) as demo: