prepare_data.py

"""
Data preparation script for downloading and preparing finetuning data:
- New language data from FineWeb2-HQ (dan_Latn, fas_Arab, swe_Latn)
- English data from FineWeb-Edu (FineWeb2-HQ does not include English)
"""

import argparse
import json
from pathlib import Path

from datasets import load_dataset
from tqdm import tqdm


def download_new_language_data(
    language: str,
    output_dir: Path,
    num_samples: int | None = None,
    full_download: bool = False,
    streaming: bool = True,
):
    """
    Download new language data from FineWeb2-HQ.


    Args:
        language: Language code (e.g. dan_Latn, fas_Arab, swe_Latn).
        output_dir: Directory where the output JSONL file will be written.
        num_samples: Number of samples to collect when full_download is False.
        full_download: If True, download the entire split; num_samples is ignored.
        streaming: Whether to stream the dataset instead of loading it into memory.
    """
    print(f"\n{'=' * 60}")
    print(f"Downloading new language from FineWeb2-HQ: {language}")
    if full_download:
        print("Mode: full download (entire train split)")
    else:
        print(f"Mode: sampled download (target: {num_samples} samples)")
    print(f"{'=' * 60}\n")

    # Load the language-specific subset from FineWeb2-HQ in streaming mode.
    try:
        dataset = load_dataset(
            "HuggingFaceFW/fineweb-2-HQ",
            name=language,
            split="train",
            streaming=streaming,
        )
        print(f"Successfully loaded dataset: {language}")
    except Exception as e:
        print(f"Failed to load dataset: {e}")
        return

    output_dir.mkdir(parents=True, exist_ok=True)
    output_file = output_dir / f"{language}.jsonl"

    # Write samples as JSONL (one JSON object per line).
    progress_total = None if full_download else num_samples
    print("Writing samples to disk...")
    with open(output_file, "w", encoding="utf-8") as f:
        for i, sample in enumerate(tqdm(dataset, total=progress_total)):
            # Stop early when a fixed sample count is requested.
            if not full_download and i >= num_samples:
                break
            json.dump(sample, f, ensure_ascii=False)
            f.write("\n")

    print(f"Data saved to: {output_file}")
    print(f"  File size: {output_file.stat().st_size / 1024 / 1024:.2f} MB\n")


def download_english_data(
    output_dir: Path,
    num_samples: int | None = None,
    full_download: bool = False,
    subset: str = "sample-10BT",
    streaming: bool = True,
):
    """
    Download English data from FineWeb-Edu.

    FineWeb2-HQ does not include English (only non-English multilingual data),
    so English data is sourced from FineWeb-Edu instead.

    Args:
        output_dir: Directory where the output JSONL file will be written.
        num_samples: Number of samples to collect when full_download is False.
        full_download: If True, download the entire split; num_samples is ignored.
        subset: FineWeb-Edu subset (sample-10BT, sample-100BT, sample-350BT, default).
        streaming: Whether to stream the dataset instead of loading it into memory.
    """
    print(f"\n{'=' * 60}")
    print("Downloading English data from FineWeb-Edu")
    print(f"  Subset: {subset}")
    if full_download:
        print("  Mode: full download (entire train split)")
    else:
        print(f"  Mode: sampled download (target: {num_samples} samples)")
    print("  Note: FineWeb2-HQ has no English; using FineWeb-Edu instead")
    print(f"{'=' * 60}\n")

    # Load the requested FineWeb-Edu subset in streaming mode.
    try:
        dataset = load_dataset(
            "HuggingFaceFW/fineweb-edu",
            name=subset,
            split="train",
            streaming=streaming,
        )
        print(f"Successfully loaded FineWeb-Edu subset: {subset}")
    except Exception as e:
        print(f"Failed to load dataset: {e}")
        return

    output_dir.mkdir(parents=True, exist_ok=True)
    output_file = output_dir / "eng_Latn.jsonl"

    # Write samples as JSONL (one JSON object per line).
    progress_total = None if full_download else num_samples
    print("Writing samples to disk...")
    with open(output_file, "w", encoding="utf-8") as f:
        for i, sample in enumerate(tqdm(dataset, total=progress_total)):
            # Stop early when a fixed sample count is requested.
            if not full_download and i >= num_samples:
                break
            json.dump(sample, f, ensure_ascii=False)
            f.write("\n")

    print(f"Data saved to: {output_file}")
    print(f"  File size: {output_file.stat().st_size / 1024 / 1024:.2f} MB\n")


def verify_data_quality(data_file: Path, language: str):
    """
    Run basic quality checks on a downloaded JSONL file.

    Args:
        data_file: Path to the JSONL data file.
        language: Language code used for logging.
    """
    print(f"\nVerifying data quality for {language}...")

    with open(data_file, "r", encoding="utf-8") as f:
        lines = f.readlines()

    total = len(lines)
    text_lengths = []

    # Inspect the first 100 samples to estimate text length statistics.
    for line in lines[:100]:
        sample = json.loads(line)
        if "text" in sample:
            text_lengths.append(len(sample["text"]))

    if text_lengths:
        avg_length = sum(text_lengths) / len(text_lengths)
        min_length = min(text_lengths)
        max_length = max(text_lengths)

        print(f"  Total samples: {total}")
        print(f"  Average text length: {avg_length:.0f} characters")
        print(f"  Shortest text: {min_length} characters")
        print(f"  Longest text: {max_length} characters")

        # Flag potentially low-quality samples based on length heuristics.
        if min_length < 100:
            print("  Warning: some texts are very short (< 100 characters)")
        if avg_length < 500:
            print("  Warning: average text length is short (< 500 characters)")
        else:
            print("  Data quality looks good")


def main():
    parser = argparse.ArgumentParser(
        description="Prepare finetuning data for Marco-MoE"
    )
    parser.add_argument(
        "--languages",
        nargs="+",
        default=["dan_Latn", "fas_Arab", "swe_Latn"],
        help="New language codes to download from FineWeb2-HQ "
        "(default: dan_Latn fas_Arab swe_Latn)",
    )
    parser.add_argument(
        "--include-english",
        action="store_true",
        help="Also download English data from FineWeb-Edu",
    )
    parser.add_argument(
        "--full-download",
        action="store_true",
        help="Download the entire train split for each language. "
        "When set, --num-samples and --english-samples are ignored.",
    )
    parser.add_argument(
        "--num-samples",
        type=int,
        default=50000,
        help="Number of samples per new language when --full-download is not set "
        "(default: 50000)",
    )
    parser.add_argument(
        "--english-samples",
        type=int,
        default=150000,
        help="Number of English samples when --full-download is not set "
        "(default: 150000)",
    )
    parser.add_argument(
        "--english-subset",
        type=str,
        default="sample-10BT",
        choices=["sample-10BT", "sample-100BT", "sample-350BT", "default"],
        help="FineWeb-Edu subset for English data (default: sample-10BT)",
    )
    parser.add_argument(
        "--output-dir",
        type=str,
        default="../outputs/prepared_data",
        help="Output directory for JSONL files (default: ../outputs/prepared_data)",
    )
    parser.add_argument(
        "--verify",
        action="store_true",
        help="Run basic quality checks after each download",
    )

    args = parser.parse_args()
    output_dir = Path(args.output_dir)

    # Print run configuration before starting downloads.
    print("\n" + "=" * 60)
    print("Marco-MoE Finetuning Data Preparation")
    print("=" * 60)
    print("\nConfiguration:")
    print(f"  New languages (FineWeb2-HQ): {', '.join(args.languages)}")
    if args.full_download:
        print("  Download mode: full (entire train split per language)")
    else:
        print("  Download mode: sampled")
        print(f"  Samples per new language: {args.num_samples}")
    print(f"  Include English: {args.include_english}")
    if args.include_english:
        print(f"  English source: FineWeb-Edu ({args.english_subset})")
        if not args.full_download:
            print(f"  English samples: {args.english_samples}")
    print(f"  Output directory: {output_dir}")
    print()

    # Download new language data from FineWeb2-HQ.
    for lang in args.languages:
        download_new_language_data(
            language=lang,
            output_dir=output_dir,
            num_samples=args.num_samples,
            full_download=args.full_download,
            streaming=True,
        )

        if args.verify:
            verify_data_quality(output_dir / f"{lang}.jsonl", lang)

    # Download English data from FineWeb-Edu (optional).
    if args.include_english:
        download_english_data(
            output_dir=output_dir,
            num_samples=args.english_samples,
            full_download=args.full_download,
            subset=args.english_subset,
            streaming=True,
        )

        if args.verify:
            verify_data_quality(
                output_dir / "eng_Latn.jsonl", "eng_Latn (FineWeb-Edu)"
            )

    # Print a summary of all generated files.
    print("\n" + "=" * 60)
    print("Data preparation complete!")
    print("=" * 60)
    print(f"\nData location: {output_dir.absolute()}")
    print("\nGenerated files:")
    for file in sorted(output_dir.glob("*.jsonl")):
        size_mb = file.stat().st_size / 1024 / 1024
        print(f"  - {file.name}: {size_mb:.2f} MB")

    print("\nNext steps:")
    print("  1. Verify data quality: python data/prepare_data.py --verify")
    print("  2. Start finetuning:")
    print("     - Marco-Mini: bash scripts/quickstart_finetune.sh marco-mini")
    print("     - Marco-Nano: bash scripts/quickstart_finetune.sh marco-nano")


if __name__ == "__main__":
    main()