train.py

import argparse
import os
from typing import List

from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    DataCollatorForLanguageModeling,
    TrainingArguments,
    Trainer,
)

def parse_args():
    p = argparse.ArgumentParser()
    p.add_argument("--dataset", required=True, help="Path to a JSON/JSONL file with either 'text' or 'prompt'+'completion'")
    p.add_argument("--output", default="trained_model", help="Where to save the fine-tuned model")
    p.add_argument("--model_name", default="distilgpt2", help="Base model name or path")
    p.add_argument("--epochs", type=float, default=1.0)
    p.add_argument("--batch_size", type=int, default=2)
    p.add_argument("--block_size", type=int, default=256)
    return p.parse_args()

def main():
    args = parse_args()
    print("📊 Loading dataset:", args.dataset, flush=True)
    dataset = load_dataset("json", data_files=args.dataset, split="train")

    cols = dataset.column_names
    print("🧾 Columns detected:", cols, flush=True)

    print("🧠 Loading model & tokenizer:", args.model_name, flush=True)
    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token  # GPT-2 family has no pad_token by default

    model = AutoModelForCausalLM.from_pretrained(args.model_name)

    def build_texts(batch) -> List[str]:
        if "text" in batch:
            return [str(t) for t in batch["text"]]
        if "prompt" in batch and "completion" in batch:
            return [f"{str(p).rstrip()}\n{str(c)}" for p, c in zip(batch["prompt"], batch["completion"])]
        raise ValueError("Dataset must have 'text' OR ('prompt' and 'completion').")

    def tokenize(batch):
        texts = build_texts(batch)
        return tokenizer(texts, padding="max_length", truncation=True, max_length=args.block_size)

    print("🔁 Tokenizing...", flush=True)
    tokenized = dataset.map(
        tokenize,
        batched=True,
        remove_columns=cols,  # keep only tokenized fields
    )

    data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)

    print("⚙ Preparing Trainer...", flush=True)
    training_args = TrainingArguments(
        output_dir=args.output,
        overwrite_output_dir=True,
        per_device_train_batch_size=args.batch_size,
        num_train_epochs=args.epochs,
        logging_steps=5,
        save_steps=50,
        save_total_limit=1,
        report_to=[],
        gradient_accumulation_steps=1,
        fp16=False,  # CPU-friendly on Spaces
    )

    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=tokenized,
        tokenizer=tokenizer,
        data_collator=data_collator,
    )

    print("🚀 Training...", flush=True)
    trainer.train()

    print("💾 Saving model to:", args.output, flush=True)
    os.makedirs(args.output, exist_ok=True)
    trainer.save_model(args.output)
    tokenizer.save_pretrained(args.output)
    print("✅ Done.", flush=True)

if __name__ == "__main__":
    main()