app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from collections import Counter
import re

# --- CONFIGURATION ---
# REPLACE WITH YOUR USERNAME
MODEL_ID = "justhariharan/Qwen2.5-Math-1.5B-Solver"

print(f"⏳ Loading {MODEL_ID}... (CPU Mode)")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float32,
    device_map="cpu"
)

# --- HELPER FUNCTIONS ---
def extract_answer(text):
    """Extracts the number after #### or the last number found."""
    if not text: return None
    if "####" in text:
        text = text.split("####")[-1]
    pattern = r"(-?[$0-9.,]{1,})"
    matches = re.findall(pattern, text)
    return matches[-1].replace(",", "").replace("$", "").strip() if matches else None

def format_prompt(current_question, history):
    # System Prompt: Friendly Teacher
    system_prompt = """<|im_start|>system
You are a patient and friendly math teacher. 
1. Solve the problem step-by-step.
2. Explain the 'logic' simply so a student can understand.
3. Always end your final result with '#### Number'.<|im_end|>"""

    # Add History (Short Term Memory - Last 1 Turn)
    history_context = ""
    
    # --- ROBUST HISTORY CHECK (The Fix) ---
    if len(history) > 0:
        try:
            # Get the last interaction
            last_turn = history[-1]
            
            # Ensure it's a list/tuple
            if isinstance(last_turn, (list, tuple)):
                # Take only the first 2 elements (User, AI) and ignore extra metadata
                last_q = last_turn[0]
                last_a = last_turn[1]
                
                history_context = f"""
<|im_start|>user
{last_q}<|im_end|>
<|im_start|>assistant
{last_a}<|im_end|>"""
        except Exception:
            # If history format is weird, just ignore it and continue safely
            pass

    # Current Input
    user_input = f"""
<|im_start|>user
{current_question}<|im_end|>
<|im_start|>assistant
"""
    return system_prompt + history_context + user_input

def solve_single(question, history, temperature=0.6):
    """Standard generation."""
    try:
        prompt = format_prompt(question, history)
        inputs = tokenizer(prompt, return_tensors="pt")
        
        with torch.no_grad():
            outputs = model.generate(
                **inputs, 
                max_new_tokens=512,
                temperature=temperature,
                do_sample=True
            )
        
        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
        if "<|im_start|>assistant" in response:
            return response.split("<|im_start|>assistant")[-1].strip()
        return response
    except Exception as e:
        return f"Error generating response: {str(e)}"

def solve_majority_vote(question, history):
    """Smart Mode: Generates 3 answers and votes."""
    candidates = []
    raw_responses = []
    
    for _ in range(3):
        # Higher temp for variety
        resp = solve_single(question, history, temperature=0.8) 
        raw_responses.append(resp)
        ans = extract_answer(resp)
        if ans:
            candidates.append(ans)
    
    if not candidates:
        return raw_responses[0]
        
    vote_counts = Counter(candidates)
    winner_ans = vote_counts.most_common(1)[0][0]
    
    for resp in raw_responses:
        if extract_answer(resp) == winner_ans:
            return f"🏆 **High Confidence Answer (Verified 3x)**\n\n{resp}"
            
    return raw_responses[0]

# --- MAIN CHAT LOGIC ---
def chat_logic(message, history, smart_mode):
    if smart_mode:
        return solve_majority_vote(message, history)
    else:
        return solve_single(message, history)

# --- UI SETUP ---
demo = gr.ChatInterface(
    fn=chat_logic,
    additional_inputs=[
        gr.Checkbox(label="🔥 Enable Smart Mode (Slow but 82% Accurate)", value=False)
    ],
    title="🧮 AI Math Tutor (Qwen-1.5B Fine-Tuned)",
    description="""
    <b>Portfolio Project:</b> A specialized math solver fine-tuned on GSM8K using LoRA.
    <br><br>
    <a href='https://colab.research.google.com/github/HariHaran9597/Math-solver/blob/main/Try_Math_Solver.ipynb' target='_blank'>
        <img src='https://colab.research.google.com/assets/colab-badge.svg' alt='Open In Colab'/>
    </a>
    <br>
    (Click the badge above to run this on a Free GPU for faster speeds!)
    <br><br>
    <b>⚠️ Performance Note:</b> This current page runs on a slow CPU.
    """,
    examples=[
        ["If I have 30 candies and eat 12, then buy 5 more, how many do I have?", False],
        ["It takes 5 machines 5 minutes to make 5 widgets. How long for 100 machines?", True]
    ]
)

if __name__ == "__main__":
    demo.launch()