Force full reasoning chain with prefilled assistant start

app.py

@@ -10,32 +10,39 @@ client = InferenceClient(token=HF_TOKEN)
 
 MODEL = "Qwen/Qwen2.5-72B-Instruct"
 
-SYSTEM_PROMPT = """You are DeepMed-R1, a medical reasoning AI. You MUST follow this EXACT output format for every answer:
+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:
 
 <think>
-## Information Analysis
-[Extract key patient demographics, symptoms, vitals, labs, imaging findings]
+## 1. Information Analysis
+[Extract ALL key clinical features: age, sex, history, symptoms, vitals, labs, imaging]
 
-## Differential Diagnosis
-[List possible diagnoses ranked by probability]
+## 2. Differential Diagnosis  
+[List 3-5 possible diagnoses ranked by probability with brief justification]
 
-## Pathophysiology
-[Explain the disease mechanism connecting symptoms]
+## 3. Pathophysiology
+[Explain the underlying disease mechanism for the most likely diagnosis]
 
-## Evidence-Based Reasoning
-[Apply clinical criteria and guidelines]
+## 4. Evidence-Based Reasoning
+[Reference clinical criteria, guidelines, or classic presentations]
 
-## Logical Elimination
-[Evaluate each option A/B/C/D and explain why each is correct or incorrect]
+## 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>
 
-## Assessment
-[Your clinical assessment in 2-3 sentences]
+**Assessment:** [2-3 sentence clinical conclusion]
 
-## Final Answer
-\\boxed{X}
+**Final Answer:** \\boxed{X}
 
-IMPORTANT: You MUST write detailed reasoning inside <think></think> tags BEFORE giving the answer. Never skip the reasoning. Each section must have at least 2 sentences."""
+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"""
 
 
 EXAMPLES = [
@@ -53,11 +60,11 @@ def respond(message, history):
         if h[0]: messages.append({"role": "user", "content": h[0]})
         if h[1]: messages.append({"role": "assistant", "content": h[1]})
     
-    # Add user message with reinforcement to show reasoning
-    user_msg = message + "\n\nShow your complete clinical reasoning step-by-step inside <think></think> tags before answering."
-    messages.append({"role": "user", "content": user_msg})
+    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 = ""
+    response = "<think>\n## 1. Information Analysis\n"
     try:
         stream = client.chat_completion(
             model=MODEL,
@@ -74,7 +81,7 @@ def respond(message, history):
                     response += delta.content
                     yield response
     except Exception as e:
-        if not response:
+        if not response or response == "<think>\n## 1. Information Analysis\n":
             yield f"⚠️ Error: {str(e)}"
         else:
             yield response