finetune_flan_t5.py

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, TrainingArguments
from trl import SFTTrainer, DataCollatorForSeq2Seq
import torch

# Load your dataset (from the converted JSONL file)
dataset = load_dataset("json", data_files="data/med_q_n_a_converted.jsonl", split="train")

# Load tokenizer and model
model_name = "google/flan-t5-base"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

# Preprocess dataset
def preprocess(example):
    input_text = example["instruction"]
    target_text = example["output"]
    tokenized = tokenizer(
        input_text,
        max_length=512,
        truncation=True,
        padding="max_length"
    )
    with tokenizer.as_target_tokenizer():
        tokenized["labels"] = tokenizer(
            target_text,
            max_length=128,
            truncation=True,
            padding="max_length"
        )["input_ids"]
    return tokenized

tokenized_dataset = dataset.map(preprocess, remove_columns=dataset.column_names)

# Define training arguments
training_args = TrainingArguments(
    output_dir="./flan-t5-medical",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=2,
    num_train_epochs=3,
    logging_dir="./logs",
    save_strategy="epoch",
    evaluation_strategy="no",
    fp16=torch.cuda.is_available()
)

# Define data collator
data_collator = DataCollatorForSeq2Seq(tokenizer, model=model)

# Initialize trainer
trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
    tokenizer=tokenizer,
    data_collator=data_collator,
)

# Start training
trainer.train()