app.py

import os
os.system('pip install torch datasets transformers')
from transformers import GPT2LMHeadModel, GPT2Tokenizer, AdamW
from torch.utils.data import Dataset, DataLoader
from datasets import load_dataset
import torch

# Load Ultrachat dataset
dataset = load_dataset("HuggingFaceH4/ultrachat_200k", split="train_sft")

# Tokenization
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")

class MyDataset(Dataset):
    def __init__(self, data, max_length=512):
        self.data = data
        self.max_length = max_length

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

    def __getitem__(self, idx):
        # Extract relevant information from the user and assistant messages
        user_content = self.data[idx][0]['content'] if 'content' in self.data[idx][0] else ""
        assistant_content = self.data[idx][1]['content'] if 'content' in self.data[idx][1] else ""

        # Combine user and assistant messages into a single text
        text = f"User: {user_content} Assistant: {assistant_content}"

        # Tokenize the text without squeezing the tensor and convert to Long tensor
        input_ids = tokenizer.encode(text, return_tensors='pt').long()

        # Optionally truncate or pad the sequence to a maximum length
        input_ids = input_ids[:, :self.max_length]

        # If needed, pad the sequence to the max_length using torch.nn.functional.pad
        input_ids = torch.nn.functional.pad(input_ids, (0, self.max_length - input_ids.size(1)), 'constant', 0)

        return {'input_ids': input_ids}

# Create DataLoader without collate_fn
my_dataset = MyDataset(dataset['messages'])
dataloader = DataLoader(my_dataset, batch_size=4, shuffle=True)

# Load pre-trained model
model = GPT2LMHeadModel.from_pretrained("gpt2")

# Move model to GPU if available
device = torch.device("cpu")
model.to(device)

# Define optimizer
optimizer = AdamW(model.parameters(), lr=5e-5)

# Fine-tuning Loop
for epoch in range(1):
    total_loss = 0.0
    for i, batch in enumerate(dataloader):
        batch = {k: v.to(device) for k, v in batch.items()}
        outputs = model(**batch, labels=batch['input_ids'])
        loss = outputs.loss
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

        total_loss += loss.item()

        if (i + 1) % 100 == 0:  # Print loss every 100 batches
            average_loss = total_loss / 100
            print(f"Epoch: {epoch + 1}, Batch: {i + 1}, Average Loss: {average_loss:.4f}")
            total_loss = 0.0

print("Training complete!")

model.save_pretrained('/gpt2_finetuned')
tokenizer.save_pretrained('/gpt2_finetuned/tokenizer')

print("Model Saved! \n Enjoy the model Now!")