tinyllama_inference.py

import json
import re
from transformers import AutoTokenizer, AutoModelForCausalLM

# Global variables for caching the model and tokenizer
tokenizer, model = None, None

def load_model():
    global tokenizer, model
    if tokenizer is None or model is None:
        # Use the DeepSeek instruct model for code evaluation.
        model_name = "deepseek-ai/deepseek-coder-1.3b-instruct"
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForCausalLM.from_pretrained(model_name)
    return tokenizer, model

def evaluate_code(question, code):
    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:
  "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.
Question: "{question}"
Solution: "{code}"
Your response:"""
    
    tokenizer, model = load_model()
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(
        **inputs,
        max_new_tokens=100,      # Allow enough tokens for a complete response
        temperature=0.2,         # Small randomness for creativity
        pad_token_id=tokenizer.eos_token_id,
        do_sample=True            # Enable sampling to encourage generation
    )
    response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    print("Raw model response:", response_text)  # Debug output

    # Use regex to extract all JSON objects (non-greedy)
    matches = re.findall(r'\{.*?\}', response_text)
    result = None
    for m in matches:
        try:
            temp = json.loads(m)
            # Check that the parsed JSON contains both expected keys
            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."}
    
    return result

# For direct command-line testing.
if __name__ == "__main__":
    import sys
    if len(sys.argv) < 3:
        print(json.dumps({"error": "Please provide a question and code as arguments"}))
        sys.exit(1)
    question = sys.argv[1]
    code = sys.argv[2]
    result = evaluate_code(question, code)
    print(json.dumps(result))