math_expert/train.py

from transformers import TrainingArguments, Trainer
from datasets import load_dataset
import jsonlines
import os
import torch
from model import Transformer, ModelArgs
from tokenizer import Tokenizer

class MathDataset(torch.utils.data.Dataset):
    def __init__(self, tokenizer, data_paths, max_length=512):
        self.tokenizer = tokenizer
        self.max_length = max_length
        self.data = []
        
        # Load and combine data from all files
        for path in data_paths:
            with jsonlines.open(path) as reader:
                self.data.extend(list(reader))
                
    def __len__(self):
        return len(self.data)
    
    def __getitem__(self, idx):
        example = self.data[idx]
        
        # Format the input text
        if "proof_steps" in example:
            # For ProofNet-style data
            text = f"Problem: {example['problem']}\nSolution: {example['solution']}\nProof Steps:\n"
            for step in example["proof_steps"]:
                text += f"- {step['text']}\n"
        else:
            # For GSM8K-style data
            text = f"Question: {example['question']}\nAnswer: {example['answer']}"
        
        # Tokenize
        inputs = self.tokenizer(
            text,
            padding="max_length",
            truncation=True,
            max_length=self.max_length,
            return_tensors="pt"
        )
        
        # Remove batch dimension
        inputs = {k: v.squeeze(0) for k, v in inputs.items()}
        
        return {
            "input_ids": inputs["input_ids"],
            "attention_mask": inputs["attention_mask"],
            "labels": inputs["input_ids"]  # For causal LM training
        }

def main():
    # Initialize your custom model
    model_args = ModelArgs(
        dim=512,
        n_layers=8,
        n_heads=8,
        vocab_size=50000,  # Adjust based on your tokenizer
        max_seq_len=1024
    )
    model = Transformer(model_args)
    
    # Initialize your custom tokenizer
    tokenizer = Tokenizer()
    
    # Configure tokenizer
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    
    # Set up training data paths
    data_dir = os.path.join(os.path.dirname(__file__), "processed_data")
    data_paths = [
        os.path.join(data_dir, "gsm8k_processed.jsonl"),
        os.path.join(data_dir, "proofnet_processed.jsonl")
    ]
    
    # Create dataset
    dataset = MathDataset(
        tokenizer=tokenizer,
        data_paths=data_paths,
        max_length=1024  # Increased max_length for longer proofs
    )
    
    # Define training arguments
    training_args = TrainingArguments(
        output_dir="./math_expert_output",
        overwrite_output_dir=True,
        num_train_epochs=3,
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        save_steps=1000,
        save_total_limit=2,
        logging_dir="./math_expert_logs",
        logging_steps=100,
        evaluation_strategy="steps",
        eval_steps=1000,
        load_best_model_at_end=True,
        learning_rate=5e-5,
        warmup_steps=500,
        weight_decay=0.01,
        fp16=True if torch.cuda.is_available() else False
    )
    
    # Create trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=dataset,
        tokenizer=tokenizer,
    )
    
    # Start training
    print("Starting training with your custom model...")
    trainer.train()
    
    # Save the model
    output_dir = "./math_expert_model"
    os.makedirs(output_dir, exist_ok=True)
    torch.save(model.state_dict(), os.path.join(output_dir, "pytorch_model.bin"))
    model_args.save(os.path.join(output_dir, "config.json"))
    print(f"Model saved to {output_dir}")

if __name__ == "__main__":
    main()