app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

# --------- Model Config ---------
MODEL_NAME = "google/gemma-7b-it"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    quantization_config=bnb_config,
    device_map="auto"
)

# --------- General Prompt (Test-Case Based) ---------
def pre_processing(requirement, code):
    return f"""
# ROLE
You are a strict Python code evaluator.

# TASK
Determine whether the code satisfies the requirement.

# INSTRUCTIONS

Step 1: Convert the requirement into 3 clear test cases (input → expected output).
Step 2: Analyze the code logic carefully.
Step 3: Simulate the code for each test case (DO NOT execute, reason mentally).
Step 4: Compare actual vs expected outputs.

# DECISION RULE
- If ALL test cases match → YES
- If ANY test case fails → NO

# RULES
- Be strict
- Do not assume missing logic
- Do not execute code
- Do not guess
- Include edge cases (if applicable)

# OUTPUT FORMAT

Test Cases:
1. Input → Expected Output
2. Input → Expected Output
3. Input → Expected Output

Evaluation:
1. PASS/FAIL
2. PASS/FAIL
3. PASS/FAIL

Final Answer: YES or NO

# INPUT
Requirement:
{requirement}

Code:
{code}
"""

# --------- Core Function ---------
def evaluate_code(requirement, code):

    prompt = pre_processing(requirement, code)

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    outputs = model.generate(
        **inputs,
        max_new_tokens=60,
        temperature=0.1,
        do_sample=False,
        repetition_penalty=1.2
    )

    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    response_upper = response.upper()

    # Extract final answer
    if "FINAL ANSWER: YES" in response_upper:
        return "✅ YES"
    elif "FINAL ANSWER: NO" in response_upper:
        return "❌ NO"
    else:
        return "⚠️ Unable to determine\n\n" + response

# --------- UI ---------
with gr.Blocks() as app:
    gr.Markdown("## 🧠 Code Requirement Validator (Gemma 7B - General)")
    gr.Markdown("Evaluates code using test-case based reasoning (high accuracy).")

    requirement = gr.Textbox(label="Requirement", lines=4)
    code = gr.Textbox(label="Code", lines=10)

    output = gr.Textbox(label="Result")

    btn = gr.Button("Evaluate")

    btn.click(fn=evaluate_code, inputs=[requirement, code], outputs=output)

app.launch()