import os
from datasets import load_dataset
from transformers import T5ForConditionalGeneration, T5Tokenizer, Trainer, TrainingArguments

# Ensure Hugging Face cache directory is writable
os.environ["HF_HOME"] = "/app/hf_cache"
os.environ["HF_DATASETS_CACHE"] = "/app/hf_cache"
os.environ["TRANSFORMERS_CACHE"] = "/app/hf_cache"
os.makedirs("/app/t5-finetuned", exist_ok=True)
# Load dataset
dataset = load_dataset("tatsu-lab/alpaca")  # Change if using your dataset

# Check dataset structure
print("Dataset splits available:", dataset)
print("Sample row:", dataset["train"][0])

# If no 'test' split exists, create one
if "test" not in dataset:
    dataset = dataset["train"].train_test_split(test_size=0.1)

# Load tokenizer & model
model_name = "t5-large"
tokenizer = T5Tokenizer.from_pretrained(model_name)
model = T5ForConditionalGeneration.from_pretrained(model_name)

# Define tokenization function
def tokenize_function(examples):
    inputs = examples["input"]  # Ensure this matches dataset key
    targets = examples["output"]

    model_inputs = tokenizer(inputs, max_length=512, truncation=True, padding="max_length")
    labels = tokenizer(targets, max_length=512, truncation=True, padding="max_length")

    model_inputs["labels"] = labels["input_ids"]
    return model_inputs

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

#  Assign train & eval datasets
train_dataset = tokenized_datasets["train"]
eval_dataset = tokenized_datasets["test"]

print("Dataset successfully split and tokenized.")

#  Define training arguments
training_args = TrainingArguments(
    output_dir="/app/t5-finetuned",
    per_device_train_batch_size=2,  # Lowered to avoid memory issues
    per_device_eval_batch_size=2,
    num_train_epochs=1,  # Test run (increase for full fine-tuning)
    logging_steps=50,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    push_to_hub=False  # Change to True to upload the model to HF Hub
)

# Set up Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
)

# Start fine-tuning
trainer.train()

print("Fine-tuning complete!")

# Save model locally
trainer.save_model("./t5-finetuned")

print("Model saved successfully!")