| import os.path | |
| from datasets import load_from_disk, load_dataset | |
| from multiprocessing import cpu_count | |
| from glob import glob | |
| paths = glob("data/*/") | |
| for path in paths: | |
| dataset = load_dataset("parquet", data_files={"train":glob(os.path.join(path, "*.parquet"))}, streaming=False, split="train") | |
| print(dataset) | |
| dataset = dataset.map(lambda x: {"n": int(x["id"])}).sort("n") | |
| dataset = dataset.remove_columns(["n"]) | |
| for i in range(10): | |
| dataset.shard(num_shards=10, index=i).to_parquet(f"{path}/train-{i:05d}-of-{10:05d}.parquet") |