app.py

import streamlit as st
from datasets import load_dataset
from transformers import GPT2LMHeadModel, GPT2Tokenizer
import torch

ds = load_dataset("higgsfield/school-math-questions")
qa_pairs = [(item['prompt'], item['completion']) for item in ds['train']]

class MathDataset(torch.utils.data.Dataset):
    def __init__(self, qa_pairs, tokenizer, max_length=128):
        self.qa_pairs = qa_pairs
        self.tokenizer = tokenizer
        self.max_length = max_length

    def __len__(self):
        return len(self.qa_pairs)

    def __getitem__(self, idx):
        question, answer = self.qa_pairs[idx]
        input_text = f"Q: {question} A:"
        
        # Tokenize and pad input and target sequences
        input_ids = self.tokenizer.encode(input_text, truncation=True, padding="max_length", max_length=self.max_length, return_tensors="pt").squeeze(0)
        target_ids = self.tokenizer.encode(answer.strip(), truncation=True, padding="max_length", max_length=self.max_length, return_tensors="pt").squeeze(0)

        # Set the labels to -100 where input_ids are padding tokens
        target_ids[target_ids == self.tokenizer.pad_token_id] = -100
        
        return {
            "input_ids": input_ids,
            "labels": target_ids,
        }

model_name = "gpt2"
tokenizer = GPT2Tokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
model = GPT2LMHeadModel.from_pretrained(model_name)

math_dataset = MathDataset(qa_pairs, tokenizer)

from transformers import Trainer, TrainingArguments

# Set training arguments
training_args = TrainingArguments(
    output_dir="./results",
    num_train_epochs=3,
    per_device_train_batch_size=2,
    save_steps=10,
    save_total_limit=2,
)

# Create a Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=math_dataset,
)

# Fine-tune the model
trainer.train()

class MathChatBot:
    def __init__(self, model_name="gpt2"):
        self.tokenizer = GPT2Tokenizer.from_pretrained(model_name)
        self.model = GPT2LMHeadModel.from_pretrained(model_name)

    def get_response(self, question):
        input_text = f"Q: {question} A:"
        input_ids = self.tokenizer.encode(input_text, return_tensors="pt")

        output = self.model.generate(input_ids, max_length=50, num_return_sequences=1)
        answer = self.tokenizer.decode(output[0], skip_special_tokens=True)
        return answer.split("A:")[-1].strip()

# Usage
if __name__ == "__main__":
    bot = MathChatBot()
    user_input = st.text_area("Enter your question:")
    response = bot.get_response(user_input)
    st.write(f"Bot: {response}")