app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
from datasets import load_dataset, Dataset
from peft import LoraConfig, get_peft_model
import torch

# Load GPT-2 model and tokenizer
model_name = "gpt2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Custom Dataset (Improved format)
custom_data = [
    {"text": "Prompt: Who are you?\nResponse: I am Eva, a virtual voice assistant."},
    {"text": "Prompt: What is your name?\nResponse: I am Eva, how can I help you?"},
    {"text": "Prompt: What can you do?\nResponse: I can assist with answering questions, searching the web, and much more!"},
    {"text": "Prompt: Who invented the computer?\nResponse: Charles Babbage is known as the father of the computer."},
    {"text": "Prompt: Tell me a joke.\nResponse: Why don’t scientists trust atoms? Because they make up everything!"},
    {"text": "Prompt: Who is the Prime Minister of India?\nResponse: The current Prime Minister of India is Narendra Modi."},
    {"text": "Prompt: Who created you?\nResponse: I was created by an expert team specializing in AI fine-tuning and web development."},
    {"text": "Prompt: Can you introduce yourself?\nResponse: I am Eva, your AI assistant, designed to assist and provide information."}
]

# Convert custom data to a Dataset
dataset_custom = Dataset.from_list(custom_data)

def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=512)

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

# Apply LoRA for efficient fine-tuning
lora_config = LoraConfig(
    r=4,  # Reduced r for stability
    lora_alpha=16,
    lora_dropout=0.1,
    bias="none",
    target_modules=["c_attn", "c_proj"]  # LoRA targets attention layers
)

model = get_peft_model(model, lora_config)
model.gradient_checkpointing_enable()

# Training arguments
training_args = TrainingArguments(
    output_dir="gpt2_finetuned",
    auto_find_batch_size=True,
    gradient_accumulation_steps=4,
    learning_rate=3e-5,  # Lowered learning rate for improved stability
    num_train_epochs=5,  # Increased epochs for better training
    save_strategy="epoch",
    logging_dir="logs",
    bf16=True,
    push_to_hub=True
)

# Trainer setup
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets
)

trainer.train()

# Save and push the model
trainer.save_model("gpt2_finetuned")
tokenizer.save_pretrained("gpt2_finetuned")
trainer.push_to_hub()

# Gradio Interface for Responses
def generate_response(prompt):
    inputs = tokenizer(f"Prompt: {prompt}\nResponse:", return_tensors="pt")
    outputs = model.generate(**inputs, max_length=150, num_return_sequences=1, temperature=0.7, top_p=0.9)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).split("Response:")[-1].strip()

demo = gr.Interface(fn=generate_response, inputs="text", outputs="text")

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