train.py

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

def parse_args():
    ap = argparse.ArgumentParser()
    ap.add_argument("--dataset", required=True, help="JSON/JSONL file (or folder with shards)")
    ap.add_argument("--output",  default="trained_model")
    ap.add_argument("--model_name", default="Salesforce/codegen-350M-multi")
    ap.add_argument("--epochs", type=float, default=1.0)
    ap.add_argument("--batch_size", type=int, default=2)
    ap.add_argument("--block_size", type=int, default=256)
    ap.add_argument("--learning_rate", type=float, default=5e-5)
    ap.add_argument("--subset", type=int, default=0, help="Use first N rows for quick runs")
    return ap.parse_args()

def main():
    a = parse_args()
    print(f"📦 Loading dataset from: {a.dataset}", flush=True)

    if os.path.isdir(a.dataset):
        pattern = os.path.join(a.dataset, "/.jsonl")
        ds = load_dataset("json", data_files=pattern, split="train")
    else:
        ds = load_dataset("json", data_files=a.dataset, split="train")

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

    # Accept either {"text": "..."} or {"prompt": "...", "completion": "..."}
    def to_text(example):
        if "text" in example:
            return example["text"]
        if "prompt" in example and "completion" in example:
            return (str(example["prompt"]).rstrip() + "\n" + str(example["completion"]))
        raise ValueError("Dataset must have 'text' or 'prompt' + 'completion'.")

    if a.subset and a.subset > 0:
        ds = ds.select(range(min(a.subset, len(ds))))
        print(f"✂ Subset: {len(ds)} rows", flush=True)

    print(f"🧠 Loading model: {a.model_name}", flush=True)
    tok = AutoTokenizer.from_pretrained(a.model_name, use_fast=True)
    if tok.pad_token is None and tok.eos_token is not None:
        tok.pad_token = tok.eos_token
    model = AutoModelForCausalLM.from_pretrained(a.model_name)

    def tokenize(batch):
        texts = [to_text(x) for x in batch]
        return tok(texts, padding="max_length", truncation=True, max_length=a.block_size)

    print("🔁 Tokenizing…", flush=True)
    tokenized = ds.map(tokenize, batched=True, remove_columns=cols)
    collator = DataCollatorForLanguageModeling(tokenizer=tok, mlm=False)

    args = TrainingArguments(
        output_dir=a.output,
        overwrite_output_dir=True,
        per_device_train_batch_size=a.batch_size,
        num_train_epochs=a.epochs,
        learning_rate=a.learning_rate,
        logging_steps=5,
        save_steps=200,
        save_total_limit=1,
        report_to=[],
        fp16=False,  # CPU-friendly in Spaces
    )

    print("⚙ Trainer…", flush=True)
    trainer = Trainer(model=model, args=args, train_dataset=tokenized,
                      tokenizer=tok, data_collator=collator)

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

    print(f"💾 Saving to {a.output}", flush=True)
    os.makedirs(a.output, exist_ok=True)
    trainer.save_model(a.output)
    tok.save_pretrained(a.output)
    print("✅ Done.", flush=True)

if __name__ == "__main__":
    try:
        main()
    except Exception as e:
        print(f"❌ Error during training: {e}", flush=True)
        raise