train.py

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"

# Load dataset
dataset = load_dataset("tatsu-lab/alpaca")
dataset["train"] = dataset["train"].select(range(2000))

# 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)
model.gradient_checkpointing_disable() 

# 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="/tmp/results",  # Use /tmp/ to avoid permission errors
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    evaluation_strategy="steps",
    save_steps=500,
    eval_steps=500,
    logging_dir="/tmp/logs",  # Avoid writing to restricted directories
    logging_steps=100,
    save_total_limit=2,
    fp16=True
)
# Set up Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
)

save_dir = "/tmp/t5-finetuned"  # Use /tmp/, which is writable
os.makedirs(save_dir, exist_ok=True)  # Ensure the directory exists
trainer.save_model(save_dir)  # Save the model

# Start fine-tuning
trainer.train()

print("Fine-tuning complete!")

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

print("Model saved successfully!")