# app.py

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

# -------------------------
# 1. Load Dataset
# -------------------------
print("Loading dataset...")
dataset = load_dataset("json", data_files={
    "train": [
        "harmless-base/train.jsonl.gz",
        "helpful-base/train.jsonl.gz",
        "helpful-online/train.jsonl.gz",
        "helpful-rejection-sampled/train.jsonl.gz"
    ],
    "test": [
        "harmless-base/test.jsonl.gz",
        "helpful-base/test.jsonl.gz",
        "helpful-online/test.jsonl.gz",
        "helpful-rejection-sampled/test.jsonl.gz"
    ]
})

# -------------------------
# 2. Load Model & Tokenizer
# -------------------------
model_name = "distilgpt2"
print(f"Loading model: {model_name}")
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Ensure tokenizer pads if needed
tokenizer.pad_token = tokenizer.eos_token

# -------------------------
# 3. Tokenize Dataset
# -------------------------
def tokenize_function(example):
    # Combine 'input' and 'output' if they exist
    text = example.get("input", "") + "\n" + example.get("output", "")
    return tokenizer(text, truncation=True, padding="max_length", max_length=128)

print("Tokenizing dataset...")
tokenized_datasets = dataset.map(tokenize_function, batched=True)

# -------------------------
# 4. Data Collator
# -------------------------
data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer, mlm=False
)

# -------------------------
# 5. Training Arguments
# -------------------------
training_args = TrainingArguments(
    output_dir="./trained_model",
    overwrite_output_dir=True,
    num_train_epochs=1,
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    save_steps=500,
    save_total_limit=2,
    logging_steps=50,
    evaluation_strategy="steps",
    eval_steps=200,
    learning_rate=5e-5,
    fp16=torch.cuda.is_available(),
    push_to_hub=False
)

# -------------------------
# 6. Trainer
# -------------------------
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["test"],
    tokenizer=tokenizer,
    data_collator=data_collator
)

# -------------------------
# 7. Train the Model
# -------------------------
print("Starting training...")
trainer.train()
print("Training complete! Saving model...")
trainer.save_model("./trained_model")
tokenizer.save_pretrained("./trained_model")

# -------------------------
# 8. Launch Gradio Interface
# -------------------------
print("Launching Gradio interface...")

def generate_response(prompt):
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(
        **inputs, 
        max_length=150, 
        do_sample=True, 
        top_p=0.95, 
        top_k=50
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

gr.Interface(
    fn=generate_response,
    inputs="text",
    outputs="text",
    title="Fine-tuned GPT-2 Chat",
    description="Ask the fine-tuned GPT-2 model anything!"
).launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))