data/opensubtitles/create.py

# /// script
# requires-python = "==3.12"
# dependencies = [
#     "datasets==3.2.0",
#     "spacy==3.8.3",
# ]
# ///

from pathlib import Path
from typing import cast

import pandas as pd
import spacy
from datasets import Dataset, load_dataset

# KCE: mail from Leon
sample_to_redact = {
    # Der kommer en dag
    "opensub_6726481",
    "opensub_6732371",
    # Kollektivet
    "opensub_6645818",
    # Flaskepost fra P
    "opensub_6666922",
    "opensub_6720216",
    "opensub_6958711",
    # Fasandræberne
    "opensub_6036947",
    "opensub_6008622",
    # En du elsker
    "opensub_5828376",
    "opensub_5828378",
    # En chance til
    "opensub_6177523",
    # Lev stærkt
    "opensub_6467655",
    # Nymphomaniac
    "opensub_5604391",
    "opensub_5748340",
    "opensub_5748494",
    "opensub_5629516",
    # Kvinden i buret
    "opensub_5636248",
    "opensub_5514603",
    "opensub_5504932",
    # Den skaldede frisør
    "opensub_5084880",
    "opensub_5031826",
    # Jagten
    "opensub_6929419",
    "opensub_4885548",
    # Melancholia
    "opensub_4421330",
    "opensub_4406991",
    "opensub_4418817",
    # Ambassadøren
    "opensub_4557721",
    # Antichrist
    "opensub_5511502",
    "opensub_3938655",
    "opensub_3636940",
    "opensub_3564521",
    "opensub_3562215",
    # En kongelig affære
    "opensub_4725493",
    "opensub_4725160",
    "opensub_4725159",
    "opensub_4916871",
    "opensub_5186746",
    # Brødre
    "opensub_233943",
    "opensub_87475",
}

column_order = [
    "text",
    "source",
    "id",
    "added",
    "created",
    "license",
    "domain",
    "metadata",
]


def convert_sample(example: dict) -> dict:
    text = example["text"]
    if example["doc_id"] in sample_to_redact:
        nlp = spacy.blank("da")
        doc = nlp(text)
        text = doc[:200].text  # first 200 words

    new_example = dict(
        text_new=text,
        id=example["doc_id"],
        source="opensubtitles",
        domain="Conversation",
        license="Creative Commons Legal Code\n\nCC0 1.0 Universal",
        added="2025-01-02",
        created="1920-01-01, 2018-01-01",  # assuming v2018
        metadata={"source-pretty": "OpenSubtitles"},
    )

    return new_example


def main():
    ds = load_dataset("DDSC/partial-danish-gigaword-no-twitter", split="train")
    ds = cast(Dataset, ds)
    ds = ds.filter(lambda x: x["source"] == "opensub", num_proc=4)
    ds = ds.map(convert_sample, num_proc=4)
    ds = ds.select_columns(column_order[1:] + ["text_new"])
    ds = ds.rename_columns({"text_new": "text"})
    # ensure order
    ds = ds.select_columns(column_order)

    df = ds.to_pandas()
    df = cast(pd.DataFrame, df)
    dedup_df = df.drop_duplicates(keep="first", subset=["text"])
    print("N. duplicates: ", df.shape[0] - dedup_df.shape[0])  # 2422

    ds = ds.select(dedup_df.index)
    assert len(set(ds["text"])) == len(ds)

    save_path = Path(__file__).parent / "opensubtitles.parquet"
    ds.to_parquet(save_path)


if __name__ == "__main__":
    main()