app.py

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

MODEL_NAME = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
DATASET_NAME = "embedding-data/Amazon-QA"
FINETUNED_MODEL_NAME = "tinyllama-shopify-lora"

def load_data():
    dataset = load_dataset(DATASET_NAME)
    df = dataset['train'].to_pandas()
    df = df.rename(columns={'query': 'question', 'pos': 'answer'})[['question', 'answer']]
    df = df.dropna().astype(str)
    df['answer'] = df['answer'].str.replace(r'\[\^.*?\^\]', '').str.strip()
    return Dataset.from_pandas(df).train_test_split(test_size=0.1)  # Now using imported Dataset

def setup_model():
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
    tokenizer.pad_token = tokenizer.eos_token

    model = AutoModelForCausalLM.from_pretrained(
        MODEL_NAME,
        device_map="auto",
        torch_dtype=torch.float32
    )

    peft_config = LoraConfig(
        r=8,
        lora_alpha=16,
        target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
        lora_dropout=0.05,
        bias="none",
        task_type="CAUSAL_LM"
    )
    return tokenizer, get_peft_model(model, peft_config)

def train_model():
    tokenizer, model = setup_model()
    data = load_data()

    def tokenize_function(examples):
        text = [f"Question: {q}\nAnswer: {a}" for q, a in zip(examples["question"], examples["answer"])]
        return tokenizer(text, truncation=True, max_length=256, padding="max_length")

    tokenized_data = data.map(tokenize_function, batched=True, remove_columns=["question", "answer"])

    trainer = Trainer(
        model=model,
        args=TrainingArguments(
            output_dir="./results",
            per_device_train_batch_size=2,
            num_train_epochs=1,
            learning_rate=2e-5,
            logging_steps=10,
            save_strategy="steps",
            save_steps=100,
            optim="adamw_torch",
            no_cuda=True
        ),
        train_dataset=tokenized_data["train"],
        data_collator=DataCollatorForLanguageModeling(tokenizer, mlm=False),
    )

    print("Starting training...")
    trainer.train()
    model.save_pretrained(FINETUNED_MODEL_NAME)
    tokenizer.save_pretrained(FINETUNED_MODEL_NAME)
    print("Training complete!")

def generate_response(message, history):
    tokenizer = AutoTokenizer.from_pretrained(FINETUNED_MODEL_NAME)
    model = AutoModelForCausalLM.from_pretrained(FINETUNED_MODEL_NAME, torch_dtype=torch.float32)

    prompt = f"Question: {message}\nAnswer:"
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=64)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).split("Answer:")[-1]

def create_interface():
    if not os.path.exists(FINETUNED_MODEL_NAME):
        print("Model not found. Starting training...")
        train_model()

    return gr.ChatInterface(
        fn=generate_response,
        examples=[
            "What's your return policy?",
            "Do you ship internationally?",
            "Is this waterproof?"
        ]
    )

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