Refine evaluation prompt and extraction for DeepSeek instruct model

tinyllama_inference.py

@@ -15,7 +15,7 @@ def load_model():
     return tokenizer, model
 
 def evaluate_code(question, code):
-    # Refined prompt with explicit rating criteria.
+    # Refined prompt with explicit instructions and a "Your response:" line.
     prompt = f"""You are an expert code evaluator.
 Evaluate the following solution for the given problem.
 Rate the solution as follows:
@@ -38,27 +38,24 @@ Your response:"""
     inputs = tokenizer(prompt, return_tensors="pt")
     outputs = model.generate(
         **inputs,
-        max_new_tokens=120,      # Increase token allowance for a complete evaluation
-        temperature=0.1,         # A low temperature for more deterministic output
+        max_new_tokens=100,      # Increase token allowance if needed
+        temperature=0.2,         # Allow some creativity, but mostly deterministic
         pad_token_id=tokenizer.eos_token_id,
-        do_sample=True            # Enable sampling to allow some creativity
+        do_sample=True
     )
     response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
-    print("Raw model response:", response_text)  # Debug: inspect raw output
+    print("Raw model response:", response_text)  # Debug output
 
-    # Extract all JSON objects using non-greedy regex and select the one with expected keys
+    # Extract all JSON objects (non-greedy) and use the last one
     matches = re.findall(r'\{.*?\}', response_text)
-    result = None
-    for m in matches:
+    if matches:
+        json_text = matches[-1]  # Pick the last JSON block
         try:
-            temp = json.loads(m)
-            if isinstance(temp, dict) and "stars" in temp and "feedback" in temp:
-                result = temp
-                break
-        except Exception:
-            continue
-    if result is None:
-        result = {"stars": 0, "feedback": "Evaluation failed. Unable to extract valid JSON from AI response."}
+            result = json.loads(json_text)
+        except Exception as e:
+            result = {"stars": 0, "feedback": "Evaluation failed. Unable to parse AI response."}
+    else:
+        result = {"stars": 0, "feedback": "Evaluation failed. Unable to extract JSON from AI response."}
     
     return result