scripts/pretrain_base_datasets.py

pretrain_base_datasets = [
    #
    # multilingual
    #
    # 3.17 GB, 2,226,907
    *[
        {'kind': 'base', 'path': 'ontocord/fineweb-permissive-multilingual-2m', 'split': f'train[{i}%:{i + 10}%]', 'format': lambda n: n['text']}
        for i in range(0, 100, 10)
    ],
    # 1.64 GB, 1,001,000
    *[
        {'kind': 'base', 'path': 'distily/c4_multilingual_1M', 'split': f'train[{i}%:{i + 10}%]', 'format': lambda n: n['text']}
        for i in range(0, 100, 10)
    ],
    # 3.8 GB, 19,454,996
    *[
        {'kind': 'base', 'path': 'sentence-transformers/parallel-sentences-wikimatrix', 'data_dir': 'all', 'split': f'train[{i}%:{i + 5}%]', 'format': lambda n: n['non_english']}
        for i in range(0, 100, 5)
    ],
    # 193 MB, 1,141,967
    *[
        {'kind': 'base', 'path': 'xu-song/cc100-samples', 'name': name, 'split': 'train', 'format': lambda n: n['text']}
        for name in [
            'am', 'ar', 'as', 'az', 'be', 'bg', 'bn', 'bn_rom', 'br',
            'bs', 'ca', 'cs', 'cy', 'da', 'de', 'el', 'en', 'eo', 'es',
            'et', 'eu', 'fa', 'ff', 'fi', 'fr', 'fy', 'ga', 'gd', 'gl',
            'gn', 'gu', 'ha', 'he', 'hi', 'hi_rom', 'hr', 'ht', 'hu',
            'hy', 'id', 'ig', 'is', 'it', 'ja', 'jv', 'ka', 'kk', 'km',
            'kn', 'ko', 'ku', 'ky', 'la', 'lg', 'li', 'ln', 'lo', 'lt',
            'lv', 'mg', 'mk', 'ml', 'mn', 'mr', 'ms', 'my', 'my_zaw',
            'ne', 'nl', 'no', 'ns', 'om', 'or', 'pa', 'pl', 'ps', 'pt',
            'qu', 'rm', 'ro', 'ru', 'sa', 'si', 'sc', 'sd', 'sk', 'sl',
            'so', 'sq', 'sr', 'ss', 'su', 'sv', 'sw', 'ta', 'ta_rom',
            'te', 'te_rom', 'th', 'tl', 'tn', 'tr', 'ug', 'uk', 'ur',
            'ur_rom', 'uz', 'vi', 'wo', 'xh', 'yi', 'yo',
            'zh-Hans', 'zh-Hant', 'zu',
        ]
    ],

    #
    # general knowledge
    #
    # 65.1 MB, 7,819
    {'kind': 'base', 'path': 'Sketched33/Cities_Wikipedia_Information', 'format': lambda n: n['wikipedia_content']},
    # 135 MB, 1,795
    {'kind': 'base', 'path': 'open-phi/textbooks', 'format': lambda n: n['markdown']},
    # 631 MB, 111,048
    {'kind': 'base', 'path': 'open-phi/programming_books_llama', 'format': lambda n: n['markdown']},

    #
    # misc
    #
    # 472 KB, 5,034
    {'kind': 'base', 'path': 'badrex/llm-emoji-dataset', 'format': '{short description}. {LLM description}. {character}'},

    #
    # math
    #
    # 12.6 GB, 14M rows
    *[
        {'kind': 'base', 'path': 'nvidia/OpenMathInstruct-2', 'split': f'train[{i}%:{i + 5}%]', 'format': '{problem} {generated_solution} {expected_answer}'}
        for i in range(0, 100, 5)
    ],

    #
    # stem
    #
    # 1.44 GB, 63,357
    *[
        {'kind': 'base', 'path': 'neuralwork/arxiver', 'split': f'train[{i}%:{i + 10}%]', 'format': lambda n: n['abstract']}
        for i in range(0, 100, 10)
    ],
    *[
        {'kind': 'base', 'path': 'neuralwork/arxiver', 'split': f'train[{i}%:{i + 10}%]', 'format': lambda n: n['markdown']}
        for i in range(0, 100, 10)
    ],

    #
    # code
    #
    # 7.81 GB, ~2,804,025
    *[
        {'kind': 'base', 'path': 'rombodawg/code_bagel_hermes-2.5', 'split': f'train[{i}%:{i + 10}%]', 'format': '{input} {output}'}
        for i in range(0, 100, 10)
    ],

    #
    # general knowledge
    #
    # 3.18 GB, 1,010,500 - paper says that extracted is 6GB
    *[
        {'kind': 'base', 'path': 'JeanKaddour/minipile', 'split': f'train[{i}%:{i + 10}%]', 'format': lambda n: n['text']}
        for i in range(0, 100, 10)
    ],
    {'kind': 'base', 'path': 'JeanKaddour/minipile', 'split': 'validation', 'format': lambda n: n['text']},
    {'kind': 'base', 'path': 'JeanKaddour/minipile', 'split': 'test', 'format': lambda n: n['text']},
]