Update tinyllama_inference.py to use deepseek-ai/deepseek-coder-1.3b-instruct

tinyllama_inference.py

@@ -15,31 +15,29 @@ def load_model():
     return tokenizer, model
 
 def evaluate_code(question, code):
-    # Updated prompt: instructs the model to output exactly valid JSON
+    # Refined prompt: clearly instructs the model to output exactly one JSON object.
     prompt = f"""You are an expert code evaluator.
-Evaluate the following solution for the given problem.
-Respond with exactly one JSON object (with no extra text) that has exactly two keys:
+Evaluate the following solution and provide your evaluation as a valid JSON object.
+The JSON object must have exactly two keys:
   "stars": an integer between 0 and 5 (0 means completely incorrect, 5 means excellent),
-  "feedback": a concise string message.
-The JSON must start with '{{' and end with '}}'.
-Do not output anything else.
-Problem: "{question}"
+  "feedback": a concise string message explaining your evaluation.
+Do not output any text besides the JSON.
+Question: "{question}"
 Solution: "{code}"
-"""
+Your response:"""
     tokenizer, model = load_model()
     inputs = tokenizer(prompt, return_tensors="pt")
-    # Adjust parameters for concise and deterministic output
     outputs = model.generate(
         **inputs,
-        max_new_tokens=60,     # Limit output length for faster responses
-        temperature=0.0,       # Deterministic output
+        max_new_tokens=100,      # Allow enough tokens for a complete response
+        temperature=0.2,         # Small randomness for creativity but mostly deterministic
         pad_token_id=tokenizer.eos_token_id,
-        do_sample=False
+        do_sample=True            # Enable sampling to encourage model generation
     )
     response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
-    print("Raw model response:", response_text)  # Debug: Inspect raw output
-
-    # Use regex (non-greedy) to extract the first JSON object from the response
+    print("Raw model response:", response_text)  # Debug: view raw output
+    
+    # Use non-greedy regex to extract the JSON object
     match = re.search(r'\{.*?\}', response_text)
     if match:
         json_text = match.group(0)