latin-llm-streaming.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "unsloth",
#     "datasets",
#     "trl",
#     "huggingface_hub",
#     "wandb",
# ]
# ///
"""
Train an LLM on Latin using streaming datasets.

Demonstrates continued pretraining with streaming - no disk space needed.
Uses FineWeb-2's Latin subset (1.47M texts, ~1.7GB).

Run locally (if you have a GPU):
    uv run latin-llm-streaming.py

Run on HF Jobs:
    hf jobs uv run latin-llm-streaming.py --flavor a100-large --secrets HF_TOKEN

With custom settings:
    hf jobs uv run latin-llm-streaming.py --flavor a100-large --secrets HF_TOKEN -- \
        --max-steps 1000 --output-repo your-username/qwen-latin
"""

import argparse
import time
import os


def parse_args():
    parser = argparse.ArgumentParser(
        description="Train an LLM on Latin using streaming datasets"
    )
    parser.add_argument(
        "--base-model",
        default="unsloth/Qwen3-0.6B-Base-unsloth-bnb-4bit",
        help="Base model to fine-tune",
    )
    parser.add_argument(
        "--output-repo",
        default=None,
        help="HF Hub repo to push model to (e.g., 'username/qwen-latin')",
    )
    parser.add_argument(
        "--max-steps",
        type=int,
        default=500,
        help="Number of training steps (default: 500)",
    )
    parser.add_argument(
        "--batch-size",
        type=int,
        default=4,
        help="Per-device batch size (default: 4)",
    )
    parser.add_argument(
        "--gradient-accumulation",
        type=int,
        default=4,
        help="Gradient accumulation steps (default: 4)",
    )
    parser.add_argument(
        "--learning-rate",
        type=float,
        default=2e-4,
        help="Learning rate (default: 2e-4)",
    )
    parser.add_argument(
        "--max-seq-length",
        type=int,
        default=2048,
        help="Maximum sequence length (default: 2048)",
    )
    parser.add_argument(
        "--lora-r",
        type=int,
        default=16,
        help="LoRA rank (default: 16)",
    )
    parser.add_argument(
        "--save-local",
        default="latin-llm-output",
        help="Local directory to save model (default: latin-llm-output)",
    )
    parser.add_argument(
        "--wandb-project",
        default="latin-llm-streaming",
        help="Wandb project name (default: latin-llm-streaming)",
    )
    parser.add_argument(
        "--wandb-run-name",
        default=None,
        help="Wandb run name (default: auto-generated)",
    )
    return parser.parse_args()


def main():
    args = parse_args()

    print("=" * 70)
    print("Latin LLM Training with Streaming Datasets")
    print("=" * 70)
    print(f"\nConfiguration:")
    print(f"  Base model:      {args.base_model}")
    print(f"  Max steps:       {args.max_steps}")
    print(f"  Batch size:      {args.batch_size} x {args.gradient_accumulation} = {args.batch_size * args.gradient_accumulation}")
    print(f"  Learning rate:   {args.learning_rate}")
    print(f"  LoRA rank:       {args.lora_r}")
    print(f"  Output repo:     {args.output_repo or '(local only)'}")
    print(f"  Wandb project:   {args.wandb_project}")
    print()

    # Import here to show progress
    from unsloth import FastLanguageModel
    from datasets import load_dataset
    from trl import SFTTrainer, SFTConfig
    from huggingface_hub import login

    # Login if pushing to hub
    if args.output_repo:
        token = os.environ.get("HF_TOKEN")
        if token:
            login(token=token)
            print("✓ Logged in to Hugging Face Hub")
        else:
            print("⚠ HF_TOKEN not set - model will only be saved locally")
            args.output_repo = None

    # Initialize wandb
    import wandb
    wandb_key = os.environ.get("WANDB_API_KEY")
    if wandb_key:
        wandb.login(key=wandb_key)
    wandb.init(
        project=args.wandb_project,
        name=args.wandb_run_name or f"latin-{args.max_steps}steps",
        config={
            "base_model": args.base_model,
            "max_steps": args.max_steps,
            "batch_size": args.batch_size,
            "gradient_accumulation": args.gradient_accumulation,
            "effective_batch_size": args.batch_size * args.gradient_accumulation,
            "learning_rate": args.learning_rate,
            "lora_r": args.lora_r,
            "max_seq_length": args.max_seq_length,
            "dataset": "HuggingFaceFW/fineweb-2 (lat_Latn)",
        },
    )
    print(f"✓ Wandb initialized: {wandb.run.url}")

    # 1. Load model
    print("\n[1/5] Loading model...")
    start = time.time()

    model, tokenizer = FastLanguageModel.from_pretrained(
        args.base_model,
        max_seq_length=args.max_seq_length,
        load_in_4bit=True,
    )

    model = FastLanguageModel.get_peft_model(
        model,
        r=args.lora_r,
        lora_alpha=args.lora_r * 2,
        lora_dropout=0,
        target_modules=[
            "q_proj", "k_proj", "v_proj", "o_proj",
            "gate_proj", "up_proj", "down_proj"
        ],
        bias="none",
        use_gradient_checkpointing="unsloth",
        random_state=3407,
    )
    print(f"✓ Model loaded in {time.time() - start:.1f}s")

    # 2. Load streaming dataset
    print("\n[2/5] Loading streaming dataset (FineWeb-2 Latin)...")
    start = time.time()

    dataset = load_dataset(
        "HuggingFaceFW/fineweb-2",
        name="lat_Latn",
        split="train",
        streaming=True,
    )

    # Peek at the data
    sample = next(iter(dataset))
    print(f"✓ Dataset ready in {time.time() - start:.1f}s")
    print(f"  Sample: {sample['text'][:100]}...")

    # 3. Format dataset
    print("\n[3/5] Preparing dataset...")

    def format_text(example):
        return {"text": example["text"] + tokenizer.eos_token}

    formatted_dataset = dataset.map(format_text)

    # 4. Train
    print(f"\n[4/5] Training for {args.max_steps} steps...")
    start = time.time()

    trainer = SFTTrainer(
        model=model,
        tokenizer=tokenizer,
        train_dataset=formatted_dataset,
        args=SFTConfig(
            per_device_train_batch_size=args.batch_size,
            gradient_accumulation_steps=args.gradient_accumulation,
            warmup_steps=min(10, args.max_steps // 10),
            max_steps=args.max_steps,
            learning_rate=args.learning_rate,
            logging_steps=max(1, args.max_steps // 20),
            optim="adamw_8bit",
            weight_decay=0.01,
            lr_scheduler_type="linear",
            seed=3407,
            output_dir=args.save_local,
            report_to="wandb",
            run_name=args.wandb_run_name or f"latin-{args.max_steps}steps",
            dataset_text_field="text",
            max_seq_length=args.max_seq_length,
            packing=False,
        ),
    )

    trainer.train()
    train_time = time.time() - start

    print(f"\n✓ Training completed in {train_time / 60:.1f} minutes")
    print(f"  Speed: {args.max_steps / train_time:.2f} it/s")

    # 5. Save and push
    print("\n[5/5] Saving model...")

    # Save locally
    model.save_pretrained(args.save_local)
    tokenizer.save_pretrained(args.save_local)
    print(f"✓ Saved locally to {args.save_local}/")

    # Push to hub if configured
    if args.output_repo:
        print(f"\nPushing to {args.output_repo}...")
        model.push_to_hub(args.output_repo, tokenizer=tokenizer)
        print(f"✓ Model available at: https://huggingface.co/{args.output_repo}")

    # Quick inference test
    print("\n" + "=" * 70)
    print("Quick inference test:")
    print("=" * 70)

    FastLanguageModel.for_inference(model)

    prompt = "Lingua Latina est"
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(
        **inputs,
        max_new_tokens=64,
        temperature=0.7,
        do_sample=True,
    )
    generated = tokenizer.decode(outputs[0], skip_special_tokens=True)

    print(f"\nPrompt: {prompt}")
    print(f"Generated: {generated}")

    print("\n" + "=" * 70)
    print("Done!")
    print("=" * 70)


if __name__ == "__main__":
    main()