data/cellar/create.py

# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "datasets>=3.2.0",
# ]
# ///

from pathlib import Path
from typing import cast
from datasets import Dataset, load_dataset, concatenate_datasets

source = "cellar"


def convert_sample(example):
    new_example = dict(
        text_new=example["text"],
        source=source,
        domain="Legal",
        license="cc-by-sa-4.0",
        added="2025-03-25",
        created="2024-01-01, 2026-01-01",  # Scrape happened within these years - data likely written earlier
        metadata={"source-pretty": "Cellar"},
    )

    return new_example


def main():
    data_path = Path("/work/dfm-data/pre-training/cellar/documents")
    data_paths = [p.as_posix() for p in data_path.glob("DAN*.jsonl.gz")]
    dfs = []
    for i, path in enumerate(data_paths):
        print(i, path.split("/")[-1])
        try:
            ds = load_dataset(
                "json", data_files=path, split="train"
            )  # a few datasets fail to load
            dfs.append(ds)
            print("\tSuccess")
        except Exception:
            print("\tFail")

    ds = concatenate_datasets(dsets=dfs)

    ds = cast(Dataset, ds)

    ds = ds.map(convert_sample, remove_columns=ds.column_names)
    ds = ds.rename_columns({"text_new": "text"})
    ds = ds.add_column("id", [f"{source}_{i}" for i in range(len(ds))])  # type: ignore
    ds = ds.select_columns(
        ["text", "source", "id", "added", "created", "license", "domain", "metadata"]
    )

    save_path = Path(__file__).parent / f"{source}.parquet"
    ds.to_parquet(save_path)


if __name__ == "__main__":
    main()