File size: 2,845 Bytes
e65d0ad
8206a45
 
e65d0ad
8206a45
 
3d71f2e
8206a45
 
 
4441c50
 
8206a45
 
 
e65d0ad
8206a45
4980b54
8206a45
3c44ee8
35d31e1
 
 
 
 
 
 
3c44ee8
35d31e1
 
3c44ee8
35d31e1
903f0f8
8206a45
903f0f8
b72b033
8206a45
 
 
 
4980b54
 
8206a45
4980b54
0bae633
8206a45
4980b54
b72b033
4980b54
3c44ee8
4980b54
 
8206a45
4980b54
 
 
 
 
8206a45
 
e65d0ad
8206a45
e65d0ad
8206a45
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
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):
    # 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:
  - 5 stars: Perfect solution; it is correct, efficient, and follows best practices.
  - 4 stars: Correct solution with minor issues or improvements possible.
  - 3 stars: Partially correct solution with noticeable issues.
  - 2 stars: Incorrect solution with some correct elements.
  - 1 star: Mostly incorrect solution.
  - 0 stars: Completely incorrect solution.
Respond with exactly one JSON object (with no extra text) that has exactly two keys:
  "stars": an integer between 0 and 5,
  "feedback": a concise string message explaining your rating.
The JSON must start with '{{' and end with '}}'.
Do not output any additional text.
Question: "{question}"
Solution: "{code}"
Your response:"""
    
    tokenizer, model = load_model()
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(
        **inputs,
        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
    )
    response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    print("Raw model response:", response_text)  # Debug output

    # Extract all JSON objects (non-greedy) and use the last one
    matches = re.findall(r'\{.*?\}', response_text)
    if matches:
        json_text = matches[-1]  # Pick the last JSON block
        try:
            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

# 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))