Rijgersberg/YouTube-Commons

YouTube Commons Re-upload

This is a re-upload of PleIAs' YouTube Commons, a valuable open dataset:

YouTube-Commons is a collection of audio transcripts of 2,063,066 videos shared on YouTube under a CC BY 4.0 license.

Content

The collection comprises 22,709,724 original and automatically translated transcripts from 3,156,703 videos (721,136 individual channels).

Unfortunately, there are problems with loading YouTube Commons with Hugging Face Datasets. In order to alleviate those and to further process the dataset, I took the source parquet-files and reuploaded this fixed version to HuggingFace.

Code

The code used for this reupload. It makes use of a git clone of the PleIAs/YouTube-Commons dataset.

from pathlib import Path

from datasets import load_dataset, Dataset
from tqdm import tqdm

columns = set('''video_link
video_id
title
text
channel
channel_id
date
license
original_language
language_id_method
transcription_language
source_language
word_count
character_count'''.split('\n'))

def generate():
    for filepath in tqdm(sorted(Path('/Path/To/PleIAs/YouTube-Commons').rglob('*.parquet'))):
        print(filepath)
        dataset = load_dataset("parquet",
                               data_files={'train': str(filepath)})
        for row in dataset['train']:
            keys = set(row)
            # Some of the files are missing one of these two columns.
            # Setting them to None results in an Arrow error, so we use '' instead
            if 'language_id_method' not in keys:
                row['language_id_method'] = ''
            if 'source_language' not in keys:
                row['source_language'] = ''
            if '__index_level_0__' in keys:
                del row['__index_level_0__']

            if not set(row) == columns:
                raise ValueError(f'Error in columns: {set(row)}')
            yield row

youtube = Dataset.from_generator(generate)
youtube.push_to_hub('Rijgersberg/YouTube-Commons')