scripts/train_tokenizer.py

import gc

from datasets import load_dataset
from transformers import PreTrainedTokenizerFast
from tokenizers import Tokenizer, normalizers, pre_tokenizers, processors, decoders
from tokenizers.models import BPE
from tokenizers.trainers import BpeTrainer


#
# datasets
#
def batch_iterator():
    # code
    dataset = load_dataset('bigcode/programming-languages-keywords', split='train')

    for row in dataset:
        for n in row['keywords']:
            yield n

    del dataset
    gc.collect()

    # code
    dataset = (
        load_dataset('bigcode/the-stack-smol-xs', data_dir=f'data/{name}', split='train', trust_remote_code=True)
        for name in [
            # 'batchfile' - unsafe
            # 'powershell' - unsafe
            'ada', 'agda', 'alloy', 'antlr', 'applescript', 'assembly',
            'augeas', 'awk', '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', '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 text in d['content']:
            yield text

    del dataset
    gc.collect()

    ## math - unsafe
    # dataset = load_dataset('gair-prox/open-web-math-pro', split='train[:1%]')
    #
    # for text in dataset['text']:
    #     yield text
    #
    # del dataset
    # gc.collect()

    # math
    dataset = load_dataset('OleehyO/latex-formulas', 'cleaned_formulas', split='train[:5%]')

    for text in dataset['latex_formula']:
        yield text

    del dataset
    gc.collect()

    # # text
    # dataset = load_dataset('JeanKaddour/minipile', split='train[:1%]')
    #
    # for text in dataset['text']:
    #     yield text
    #
    # del dataset
    # gc.collect()

    # text
    dataset = (
        load_dataset('saillab/taco-datasets', data_dir=data_dir, split='train[:5%]')
        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'] + '\n' + row['input'] + '\n' + row['output']

    del dataset
    gc.collect()

    # text
    dataset = (
        load_dataset('xu-song/cc100-samples', lang, split='train[:5%]')
        for lang in [
            'en', 'hr', 'sr', 'ru',
            'am', 'ar', 'as', 'az', 'be', 'bg', 'bn', 'bn_rom', 'br',
            'bs', 'ca', 'cs', 'cy', 'da', 'de', 'el', 'eo', 'es',
            'et', 'eu', 'fa', 'ff', 'fi', 'fr', 'fy', 'ga', 'gd', 'gl',
            'gn', 'gu', 'ha', 'he', 'hi', 'hi_rom', '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', 'sa', 'si', 'sc', 'sd', 'sk', 'sl',
            'so', 'sq', '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 text in d['text']:
            yield text

    del dataset
    gc.collect()


#
# special_tokens
#
special_tokens = [
    '<unk>',
    '<|begin_of_text|>',
    '<|end_of_text|>',
    '<|start_header_id|>',
    '<|end_header_id|>',
    '<|eom_id|>',
    '<|eot_id|>',
    'system',
    'user',
    'assistant',
    'tool',
    'agent',
    'internal', # thinking

    # tool/function calling
    '<tools>',
    '</tools>',
    '<tool>',
    '</tool>',
    '<tool_call>',
    '</tool_call>',
    '<tool_response>',
    '</tool_response>',
    '"name"',
    '"arguments"',

    #
    # JSON Schema
    #
    # General Metadata Keywords
    '"$schema"',
    '"$id"',
    '"$ref"',
    '"$defs"',
    '"$anchor"',
    '"$dynamicAnchor"',
    '"$dynamicRef"',
    '"$vocabulary"',
    '"$comment"',
    # Data Types
    '"null"',
    '"boolean"',
    '"object"',
    '"array"',
    '"number"',
    '"string"',
    '"integer"',
    # Validation Keywords
    '"type"',
    '"enum"',
    '"const"',
    '"multipleOf"',
    '"maximum"',
    '"exclusiveMaximum"',
    '"minimum"',
    '"exclusiveMinimum"',
    '"maxLength"',
    '"minLength"',
    '"pattern"',
    '"additionalItems"',
    '"items"',
    '"prefixItems"',
    '"contains"',
    '"maxItems"',
    '"minItems"',
    '"uniqueItems"',
    '"maxProperties"',
    '"minProperties"',
    '"required"',
    '"properties"',
    '"patternProperties"',
    '"additionalProperties"',
    '"dependentRequired"',
    '"dependentSchemas"',
    '"propertyNames"',
    # Conditional Keywords
    '"if"',
    '"then"',
    '"else"',
    '"allOf"',
    '"anyOf"',
    '"oneOf"',
    '"not"',
    # Additional Keywords for Evaluation Control
    '"unevaluatedItems"',
    '"unevaluatedProperties"',
    # Informational Keywords
    '"title"',
    '"description"',
    '"default"',
    '"deprecated"',
    '"readOnly"',
    '"writeOnly"',
    '"examples"',
    # Content-Related Keywords
    '"contentEncoding"',
    '"contentMediaType"',
    '"contentSchema"',
    # Additional Keywords
    '"next"',                # Typically used in reference to linked or next items
    '"value"',                # Represents the value of a property or item

    # misc
    '<input>',
    '</input>',
    '<output>',
    '</output>',
    '<query>',
    '</query>',
    '<key>',
    '</key>',
    '<value>',
    '</value>',
    '<text>',
    '</text>',
    '<code>',
    '</code>',
    '<image>',
    '</image>',
    '<file>',
    '</file>',

    # qa
    '<question>',
    '</question>',
    '<answer>',
    '</answer>',

    # thought
    '<thought>',
    '</thought>',
    '<plan>',
    '</plan>',
    '<vote>',
    '</vote>',
    '<passage>',
    '</passage>',

    # reasoning
    '<reasoning>',
    '</reasoning>',
    '<acting>',
    '</acting>',
    '<action>',
    '</action>',
    '<observation>',
    '</observation>',
    '<claim>',
    '</claim>',

    # reflection
    '<thinking>',
    '</thinking>',
    '<reflection>',
    '</reflection>',
    '<step>',
    '</step>',

    # graph
    '<graph>',
    '</graph>',
    '<edge>',
    '</edge>',
    '<source>',
    '</source>',
    '<destination>',
    '</destination>',
    '<relation>',
    '</relation>',
    # '<value>',
    # '</value>',
]

for i in range(2, 25):
    special_tokens.append(' ' * i)

for i in range(2, 25):
    special_tokens.append('\t' * i)

for i in range(2, 25):
    special_tokens.append('\n' * i)

for i in range(2, 25):
    special_tokens.append('\r' * i)

for i in range(2, 25):
    special_tokens.append('\r\n' * i)

for i in range(256):
    special_tokens.append(f'<0x{i:02X}>')

for i in range(256):
    special_tokens.append(f'<|reserved_special_token_{i}|>')


#
# train tokenizer
#
bpe = BPE(unk_token='<unk>', fuse_unk=True, byte_fallback=True)
tokenizer = Tokenizer(bpe)

tokenizer.normalizer = normalizers.Sequence([
    normalizers.Prepend('▁'),
    normalizers.Replace(' ', '▁'),
])

tokenizer.pre_tokenizer = None

tokenizer.post_processor = processors.TemplateProcessing(
    single='$A:0',                              # $A represents the token, :0 specifies the type ID for single sequences
    pair='$A:0 $B:1',                           # For pairs, we specify type IDs for both tokens
    special_tokens=[],
)

tokenizer.decoder = decoders.Sequence([
    decoders.Replace('▁', ' '),
    decoders.ByteFallback(),
    decoders.Fuse(),
    decoders.Strip(' ', 1, 0),
])

trainer = BpeTrainer(
    vocab_size=32768, # 32 * 1024
    min_frequency=10,
    special_tokens=special_tokens,
    max_token_length=8,
)

tokenizer.train_from_iterator(batch_iterator(), trainer)
tokenizer.save('../tokenizer.json')
tokenizer.model.save('../')

CHAT_TEMPLATE = (
    "{{ bos_token }}"

    "{% for message in messages %}"
        "{{'<|start_header_id|>' + message['role'] + '<|end_header_id|>' + message['content'] + '<|eot_id|>'}}"
    "{% endfor %}"

    "{% if add_generation_prompt %}"
        "{{ '<|start_header_id|>assistant<|end_header_id|>' }}"
    "{% else %}"
        "{{ eos_token }}"
    "{% endif %}"
)

fast_tokenizer = PreTrainedTokenizerFast(
    tokenizer_object=tokenizer,
    chat_template=CHAT_TEMPLATE,
    bos_token='<|begin_of_text|>',
    eos_token='<|end_of_text|>',
    unk_token='<unk>',
    clean_up_tokenization_spaces=True,
)

fast_tokenizer.save_pretrained('../')