split_pile_data.py

#!/usr/bin/env python3
"""
划分 Pile 数据集为 train/val/test
与 BabyLM 保持相同的命名格式
"""

import numpy as np
from pathlib import Path
import json

def split_pile_data(
    input_file="batch_0_to_1000.npy",
    output_dir="tokenized",
    train_ratio=0.8,
    val_ratio=0.1,
    seed=42
):
    """划分 Pile 数据"""
    print("="*70)
    print("📊 Splitting Pile Dataset")
    print("="*70)
    print(f"Input: {input_file}")
    print(f"Output: {output_dir}/")
    print(f"Split: train={train_ratio:.0%}, val={val_ratio:.0%}, test={1-train_ratio-val_ratio:.0%}")
    print(f"Seed: {seed}")
    print("="*70)
    print()
    
    # 加载数据
    print("📥 Loading data...")
    data = np.load(input_file, allow_pickle=False)
    print(f"  Shape: {data.shape}")
    print(f"  Dtype: {data.dtype}")
    print(f"  Total samples: {len(data):,}")
    print(f"  Sequence length: {data.shape[1]}")
    print()
    
    # 设置随机种子
    np.random.seed(seed)
    
    # 打乱索引
    print("🔀 Shuffling indices...")
    indices = np.arange(len(data))
    np.random.shuffle(indices)
    
    # 计算划分点
    n_total = len(data)
    n_train = int(n_total * train_ratio)
    n_val = int(n_total * val_ratio)
    
    # 划分索引
    train_indices = indices[:n_train]
    val_indices = indices[n_train:n_train + n_val]
    test_indices = indices[n_train + n_val:]
    
    print("✂️  Splitting...")
    print(f"  Train: {len(train_indices):,} samples ({len(train_indices)/n_total*100:.1f}%)")
    print(f"  Val:   {len(val_indices):,} samples ({len(val_indices)/n_total*100:.1f}%)")
    print(f"  Test:  {len(test_indices):,} samples ({len(test_indices)/n_total*100:.1f}%)")
    print()
    
    # 创建输出目录
    output_dir = Path(output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    
    print("💾 Saving splits...")
    
    # Train
    print(f"  → train.npy...")
    np.save(output_dir / "train.npy", data[train_indices])
    size_mb = (output_dir / "train.npy").stat().st_size / (1024**2)
    print(f"    ✅ {size_mb:.1f} MB")
    
    # Val (eval)
    print(f"  → eval.npy...")
    np.save(output_dir / "eval.npy", data[val_indices])
    size_mb = (output_dir / "eval.npy").stat().st_size / (1024**2)
    print(f"    ✅ {size_mb:.1f} MB")
    
    # Test
    print(f"  → test.npy...")
    np.save(output_dir / "test.npy", data[test_indices])
    size_mb = (output_dir / "test.npy").stat().st_size / (1024**2)
    print(f"    ✅ {size_mb:.1f} MB")
    
    # 创建 metadata.json（与 BabyLM 格式一致）
    print(f"  → metadata.json...")
    metadata = {
        "vocab_size": int(data.max()) + 1,  # 最大 token ID + 1
        "sequence_length": int(data.shape[1]),
        "num_train": int(len(train_indices)),
        "num_eval": int(len(val_indices)),
        "num_test": int(len(test_indices)),
        "total_samples": int(n_total),
        "dtype": str(data.dtype),
        "seed": seed
    }
    
    with open(output_dir / "metadata.json", 'w') as f:
        json.dump(metadata, f, indent=2)
    print(f"    ✅ Saved")
    
    print()
    print("="*70)
    print("✅ Split Complete!")
    print("="*70)
    print()
    print("📁 Output structure:")
    print(f"  {output_dir}/")
    print(f"    ├── train.npy     ({len(train_indices):,} samples)")
    print(f"    ├── eval.npy      ({len(val_indices):,} samples)")
    print(f"    ├── test.npy      ({len(test_indices):,} samples)")
    print(f"    └── metadata.json")
    print()
    print("📊 Metadata:")
    for key, value in metadata.items():
        print(f"    {key}: {value:,}" if isinstance(value, int) else f"    {key}: {value}")


if __name__ == "__main__":
    import argparse
    
    parser = argparse.ArgumentParser()
    parser.add_argument("--input", default="batch_0_to_1000.npy")
    parser.add_argument("--output", default="tokenized")
    parser.add_argument("--train_ratio", type=float, default=0.8)
    parser.add_argument("--val_ratio", type=float, default=0.1)
    parser.add_argument("--seed", type=int, default=42)
    
    args = parser.parse_args()
    
    split_pile_data(
        args.input,
        args.output,
        args.train_ratio,
        args.val_ratio,
        args.seed
    )