individual_steps/finetune_t5.py

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

# Load dataset from JSON
dataset = load_dataset('json', data_files='dataset.json')

# Split dataset into training and validation sets
train_test_split = dataset['train'].train_test_split(test_size=0.15, seed=42)
train_dataset = train_test_split['train']
eval_dataset = train_test_split['test']

# Load tokenizer
tokenizer = T5Tokenizer.from_pretrained('./t5_small_weights')

# Preprocess dataset
def preprocess_data(examples):
    inputs = ["question: " + q.strip() for q in examples['input']]
    targets = [r.strip() for r in examples['response']]
    model_inputs = tokenizer(inputs, max_length=128, truncation=True, padding='max_length')
    labels = tokenizer(targets, max_length=64, truncation=True, padding='max_length')
    model_inputs['labels'] = [
        [(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels['input_ids']
    ]
    return model_inputs

# Apply preprocessing
processed_train_dataset = train_dataset.map(preprocess_data, batched=True, remove_columns=['input', 'response'])
processed_eval_dataset = eval_dataset.map(preprocess_data, batched=True, remove_columns=['input', 'response'])

# Load model
model = T5ForConditionalGeneration.from_pretrained('./t5_small_weights')

# Define training arguments
training_args = TrainingArguments(
    output_dir='./results',
    num_train_epochs=10,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=2,
    learning_rate=3e-4,  # Slightly increased for better convergence
    save_steps=500,
    save_total_limit=2,
    logging_steps=50,
    eval_strategy="steps",
    eval_steps=100,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",
    greater_is_better=False,
    gradient_checkpointing=True,
    max_grad_norm=1.0,
)

# Initialize Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=processed_train_dataset,
    eval_dataset=processed_eval_dataset,
)

# Train the model
print("Starting training...")
trainer.train()
print("Training finished.")

# Save the fine-tuned model
final_model_save_path = './finetuned_t5_improved'
model.save_pretrained(final_model_save_path)
tokenizer.save_pretrained(final_model_save_path)
print(f"Model fine-tuned and saved to '{final_model_save_path}'")