scripts/create_splits.py

#!/usr/bin/env python3
"""
Create train/dev/test splits for UVW 2026 dataset.

UVW 2026: Underthesea Vietnamese Wikipedia Dataset
https://github.com/undertheseanlp/underthesea/issues/896

Uses streaming approach to minimize memory usage.
"""

import json
import random
from pathlib import Path

from tqdm import tqdm


DATA_DIR = Path(__file__).parent.parent / "data" / "processed"
SPLITS_DIR = Path(__file__).parent.parent / "data" / "splits"

# Input file (with wikidata and quality scores)
INPUT_FILE = DATA_DIR / "uvw_2026_wikidata.jsonl"

# Split ratios
TRAIN_RATIO = 0.8
DEV_RATIO = 0.1
TEST_RATIO = 0.1

# Random seed for reproducibility
SEED = 42


def count_lines(path: Path) -> int:
    """Count lines in file without loading into memory."""
    count = 0
    with open(path, "r", encoding="utf-8") as f:
        for _ in f:
            count += 1
    return count


def create_split_indices(total: int, seed: int = SEED) -> dict[str, list[int]]:
    """Create shuffled indices for splits without loading articles."""
    random.seed(seed)
    indices = list(range(total))
    random.shuffle(indices)

    train_end = int(total * TRAIN_RATIO)
    dev_end = train_end + int(total * DEV_RATIO)

    # Create index-to-split mapping
    index_to_split = {}
    for i, idx in enumerate(indices):
        if i < train_end:
            index_to_split[idx] = "train"
        elif i < dev_end:
            index_to_split[idx] = "dev"
        else:
            index_to_split[idx] = "test"

    return index_to_split


def main():
    """Create dataset splits using streaming."""
    if not INPUT_FILE.exists():
        print(f"Dataset not found: {INPUT_FILE}")
        print("Please run add_wikidata.py first.")
        return

    print(f"Input: {INPUT_FILE}")
    print(f"Output: {SPLITS_DIR}")

    # Count total articles
    print("\nCounting articles...")
    total = count_lines(INPUT_FILE)
    print(f"Total articles: {total:,}")

    # Create split indices
    print("\nCreating split indices...")
    index_to_split = create_split_indices(total)

    # Calculate split sizes
    split_counts = {"train": 0, "dev": 0, "test": 0}
    for split_name in index_to_split.values():
        split_counts[split_name] += 1

    for split_name, count in split_counts.items():
        pct = count / total * 100
        print(f"  {split_name}: {count:,} articles ({pct:.1f}%)")

    # Create output directory
    SPLITS_DIR.mkdir(parents=True, exist_ok=True)

    # Open all output files
    print("\nWriting splits (streaming)...")
    split_files = {
        "train": open(SPLITS_DIR / "train.jsonl", "w", encoding="utf-8"),
        "dev": open(SPLITS_DIR / "dev.jsonl", "w", encoding="utf-8"),
        "test": open(SPLITS_DIR / "test.jsonl", "w", encoding="utf-8"),
    }

    plaintext_dirs = {}
    plaintext_files = {}
    for split_name in ["train", "dev", "test"]:
        pdir = SPLITS_DIR / "plaintext" / split_name
        pdir.mkdir(parents=True, exist_ok=True)
        plaintext_dirs[split_name] = pdir
        plaintext_files[split_name] = open(pdir / "sentences.txt", "w", encoding="utf-8")

    try:
        with open(INPUT_FILE, "r", encoding="utf-8") as fin:
            for idx, line in enumerate(tqdm(fin, total=total, desc="Processing")):
                split_name = index_to_split[idx]
                article = json.loads(line)

                # Write JSONL
                split_files[split_name].write(line)

                # Write plaintext sentences
                content = article["content"]
                for sent in content.replace("\n", " ").split("."):
                    sent = sent.strip()
                    if sent and len(sent) > 10:
                        plaintext_files[split_name].write(sent + ".\n")
    finally:
        for f in split_files.values():
            f.close()
        for f in plaintext_files.values():
            f.close()

    # Save split statistics
    stats = {
        "seed": SEED,
        "source": str(INPUT_FILE.name),
        "ratios": {
            "train": TRAIN_RATIO,
            "dev": DEV_RATIO,
            "test": TEST_RATIO,
        },
        "counts": split_counts,
    }

    stats_path = SPLITS_DIR / "split_info.json"
    with open(stats_path, "w", encoding="utf-8") as f:
        json.dump(stats, f, indent=2)

    print(f"\nSplit info saved to: {stats_path}")
    print("\nOutput files:")
    for split_name in ["train", "dev", "test"]:
        print(f"  - {SPLITS_DIR / f'{split_name}.jsonl'}")
        print(f"  - {plaintext_dirs[split_name] / 'sentences.txt'}")

    print("\nSplit creation complete!")


if __name__ == "__main__":
    main()