finetune.py

import torch
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer

# Load your data
dataset = load_dataset("json", data_files={"train": "qa_data.jsonl"})

# Choose a model (GPT-2 small is easy to start)
model_name = "gpt2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Add pad token if missing (GPT-2 doesn't have one by default)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Tokenize
def preprocess(example):
    prompt = example["prompt"]
    response = example["response"]
    text = prompt + " " + response
    tokens = tokenizer(
        text,
        truncation=True,
        padding="max_length",
        max_length=128,
    )
    tokens["labels"] = tokens["input_ids"].copy()
    return tokens

tokenized = dataset["train"].map(preprocess)

# Training arguments
args = TrainingArguments(
    output_dir="gpt2-finetuned-qa",
    per_device_train_batch_size=2,
    num_train_epochs=5,
    logging_steps=10,
    save_steps=50,
    fp16=True if torch.cuda.is_available() else False,
    report_to="none",
)

# Trainer
trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized,
)

trainer.train()
model.save_pretrained("gpt2-finetuned-qa")
tokenizer.save_pretrained("gpt2-finetuned-qa")