app.py

import gradio as gr
import os
import re
import json
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, pipeline
from transformers import DataCollatorWithPadding
from huggingface_hub import login

# Retrieve the Hugging Face token from the Space secrets
token = os.getenv("HF_TOKEN")

# Log in using the token
login(token=token)

# Load the dataset
dataset = load_dataset('json', data_files='dataset.json')

# Tokenize the dataset
# Step 6: Tokenize the dataset
tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2", token=token)
tokenizer.pad_token = tokenizer.eos_token  # Set pad_token to eos_token

# Tokenize the data and ensure labels are set
def tokenize_function(examples):
    # Tokenize input text, adding labels for causal language modeling
    inputs = tokenizer(examples["text"], truncation=True, padding="max_length", max_length=256)
    
    # The labels are the input_ids shifted by one token (for causal language modeling)
    inputs["labels"] = inputs["input_ids"].copy()  # Copy the input_ids for labels
    return inputs

tokenized_datasets = dataset.map(tokenize_function, batched=True)

# Data collator
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

# Split dataset into training and validation
tokenized_datasets = tokenized_datasets['train'].train_test_split(test_size=0.1)
train_dataset = tokenized_datasets["train"]
eval_dataset = tokenized_datasets["test"]

# Fine-tune the model
model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2", token=token)

training_args = TrainingArguments(
    output_dir="./results",
    eval_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=4,  # Reduced batch size
    per_device_eval_batch_size=4,   # Reduced batch size
    num_train_epochs=3,
    weight_decay=0.01,
    report_to="none",  # Disables wandb logging
    fp16=True,  # Enable mixed precision (use 16-bit instead of 32-bit precision)
    gradient_accumulation_steps=8,  # Accumulate gradients over 8 steps
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    data_collator=data_collator
)

trainer.train()

# Save the model
model.save_pretrained("./fine-tuned-gpt2")
tokenizer.save_pretrained("./fine-tuned-gpt2")

# Evaluate the model
#results = trainer.evaluate()
#print(results)

# Create a Gradio interface for text generation
def generate_text(prompt):
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(inputs["input_ids"], max_length=50, num_return_sequences=1)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

iface = gr.Interface(fn=generate_text, inputs="text", outputs="text")
iface.launch()