scripts/prepare_pretrain_dataset.py

import gc

from datasets import load_dataset
from litdata import optimize, TokensLoader
from litgpt.tokenizer import Tokenizer
from functools import partial


def batch_iterator(name=None):
    # text
    if name in (None, 'saillab/taco-datasets'):
        dataset = (
            load_dataset(name, data_dir=data_dir, split='train')
            for data_dir in [
                'multilingual-instruction-tuning-dataset /multilingual-alpaca-52k-gpt-4',
                'multilingual-instruction-tuning-dataset /multilinugal-dolly-15k',
            ]
        )

        for d in dataset:
            for row in d:
                for n in row:
                    yield (
                        row['instruction'] +
                        ' ' +
                        row['input'] +
                        ' ' +
                        row['output']
                    )

        del dataset
        gc.collect()

    # text
    if name in (None, 'xu-song/cc100-samples'):
        dataset = (
            load_dataset(name, lang, split='train')
            for lang 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',
            ]
        )

        for d in dataset:
            for row in d['text']:
                yield row

        del dataset
        gc.collect()

    # text
    if name in (None, 'ontocord/fineweb-permissive-multilingual-2m'):
        dataset = load_dataset(name, split='train')
        
        for row in dataset['text']:
            yield row
        
        del dataset
        gc.collect()

    # text
    if name in (None, 'MuskumPillerum/General-Knowledge'):
        dataset = load_dataset(name, split='train')

        for row in dataset:
            if not row['Question'] or not row['Answer']:
                continue

            yield row['Question'] + ' ' + row['Answer']

        del dataset
        gc.collect()

    # text
    if name in (None, 'yirenc/general_knowledge_boolean'):
        for split in ['train', 'validation']:
            dataset = load_dataset(name, split=split)

            for row in dataset:
                yield row['question'] + '? ' + str(row['answer']) + '. ' + row['passage']

            del dataset
            gc.collect()

    # text
    if name in (None, 'nampdn-ai/tiny-textbooks'):
        for split in ['train', 'test']:
            dataset = load_dataset(name, split=split)

            for row in dataset['textbook']:
                yield row

            del dataset
            gc.collect()

    # code
    if name in (None, 'nampdn-ai/tiny-codes'):
        dataset = load_dataset(name, split='train')
        
        for row in dataset:
            yield (
                row['prompt'] +
                ' ' +
                row['response']
            )
        
        del dataset
        gc.collect()

    # code
    if name in (None, 'bigcode/the-stack-smol-xs'):
        dataset = (
            load_dataset(name, lang, split='train', trust_remote_code=True)
            for lang in [
                'ada', 'agda', 'alloy', 'antlr', 'applescript', 'assembly',
                'augeas', 'awk', 'batchfile', 'bison', 'bluespec', 'c',
                'c++', 'c-sharp', 'clojure', 'cmake', 'coffeescript', 'common-lisp',
                'css', 'cuda', 'dart', 'dockerfile', 'elixir',
                'elm', 'emacs-lisp','erlang', 'f-sharp', 'fortran', 'glsl', 'go',
                'groovy', 'haskell','html', 'idris', 'isabelle', 'java', 
                'java-server-pages', 'javascript', 'julia', 'kotlin', 'lean',
                'literate-agda', 'literate-coffeescript', 'literate-haskell',
                'lua', 'makefile', 'maple', 'markdown', 'mathematica', 'matlab',
                'ocaml', 'pascal', 'perl', 'php', 'powershell', 'prolog',
                'protocol-buffer', 'python', 'r', 'racket', 'restructuredtext',
                'rmarkdown', 'ruby', 'rust', 'sas', 'scala', 'scheme', 
                'shell', 'smalltalk', 'solidity', 'sparql', 'sql', 'stan',
                'standard-ml', 'stata', 'systemverilog', 'tcl', 'tcsh', 'tex', 
                'thrift', 'typescript', 'verilog', 'vhdl', 'visual-basic', 'xslt',
                'yacc', 'zig',
            ]
        )

        for d in dataset:
            for row in d['content']:
                yield row

        del dataset
        gc.collect()

    # text + code
    if name in (None, 'm-a-p/CodeFeedback-Filtered-Instruction'):
        dataset = load_dataset(name, split='train')

        for row in dataset:
            yield (
                row['query'] +
                ' ' +
                row['answer']
            )

        del dataset
        gc.collect()

    # code
    if name in (None, 'jtatman/python-code-dataset-500k'):
        dataset = load_dataset(name, split='train')

        for row in dataset:
            yield (
                row['instruction'] +
                ' ' +
                row['output']
            )

        del dataset
        gc.collect()

    # code
    if name in (None, 'iamtarun/python_code_instructions_18k_alpaca'):
        dataset = load_dataset(name, split='train')

        for row in dataset:
            yield (
                row['instruction'] +
                ' ' +
                row['input'] +
                ' ' +
                row['output']
            )

        del dataset
        gc.collect()

    # code
    if name in (None, 'HuggingFaceH4/CodeAlpaca_20K'):
        for split in ['train', 'test']:
            dataset = load_dataset(name, split=split)

            for row in dataset:
                yield (
                    row['prompt'] +
                    ' ' +
                    row['completion']
                )

            del dataset
            gc.collect()

    # math
    if name in (None, 'gair-prox/open-web-math-pro'):
        dataset = load_dataset(name, split='train')
        
        for row in dataset['text']:
            yield row
        
        del dataset
        gc.collect()

    # math
    if name in (None, 'rvv-karma/Math-QA'):
        for split in ['train', 'val', 'test']:
            dataset = load_dataset(name, split=split)
            
            for row in dataset:
                yield (
                    row['question'] +
                    ' ' +
                    row['answer']
                )
            
            del dataset
            gc.collect()

    # math
    if name in (None, 'ajibawa-2023/Maths-College'):
        dataset = load_dataset(name, split='train')
        
        for row in dataset:
            yield (
                row['instruction'] +
                ' ' +
                row['output']
            )
        
        del dataset
        gc.collect()

    # math
    if name in (None, 'microsoft/orca-math-word-problems-200k'):
        dataset = load_dataset(name, split='train')

        for row in dataset:
            yield (
                row['question'] +
                ' ' +
                row['answer']
            )

        del dataset
        gc.collect()

    # math
    if name in (None, 'fblgit/simple-math'):
        for split in ['train', 'test']:
            dataset = load_dataset(name, revision='refs/convert/parquet', split=split)

            for row in dataset:
                yield (
                    str(row['instruction']) +
                    ' = ' +
                    str(row['output'])
                )

            del dataset
            gc.collect()

    # reasoning
    if name in (None, 'SkunkworksAI/reasoning-0.01'):
        dataset = load_dataset(name, split='train')

        for row in dataset:
            yield (
                row['instruction'] +
                ' ' +
                row['reasoning'] +
                ' ' +
                row['output']
            )

        del dataset
        gc.collect()

    # emoji
    if name in (None, 'badrex/llm-emoji-dataset'):
        dataset = load_dataset(name, split='train')
        
        for row in dataset:
            yield (
                row['character'] +
                ' ' +
                row['unicode'] +
                ' ' +
                row['short description'] +
                ' ' +
                str(row['tags']) +
                ' ' +
                row['LLM description']
            )
        
        del dataset
        gc.collect()


def tokenize_fn(dataset_name, tokenizer=None):
    for text in batch_iterator(dataset_name):
        text_ids = tokenizer.encode(text, bos=False, eos=True)
        yield text_ids


datasets_names = [
    'saillab/taco-datasets',
    # 'xu-song/cc100-samples',
    # 'ontocord/fineweb-permissive-multilingual-2m',
    'MuskumPillerum/General-Knowledge',
    'yirenc/general_knowledge_boolean',
    'nampdn-ai/tiny-textbooks',
    # 'nampdn-ai/tiny-codes',
    'bigcode/the-stack-smol-xs',
    'm-a-p/CodeFeedback-Filtered-Instruction',
    # 'jtatman/python-code-dataset-500k',
    'iamtarun/python_code_instructions_18k_alpaca',
    'HuggingFaceH4/CodeAlpaca_20K',
    # 'gair-prox/open-web-math-pro',
    'rvv-karma/Math-QA',
    # 'ajibawa-2023/Maths-College',
    'microsoft/orca-math-word-problems-200k',
    'fblgit/simple-math',
    # 'SkunkworksAI/reasoning-0.01',
    'badrex/llm-emoji-dataset',
]

outputs = optimize(
    fn=partial(tokenize_fn, tokenizer=Tokenizer('..')),
    inputs=datasets_names,
    output_dir='../pretrain-data/',
    # Number of tokens to store by chunks. This is roughly 64MB of tokens per chunk.
    chunk_size=(2049 * 8012),
    num_workers=16,
)