scripts/train_tokenizer.py

import gc
import sys

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
from tokenizers.processors import TemplateProcessing


x = input('Are you sure? [y/N] ')

if x not in ('y', 'Y', 'yes'):
    sys.exit(0)


def batch_iterator():
    # text
    dataset = (
        load_dataset('saillab/taco-datasets', 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'] + '\n' + row['input'] + '\n' + row['output']

    del dataset
    gc.collect()

    # text
    dataset = (
        load_dataset('xu-song/cc100-samples', 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()

    # 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', 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:
            yield row['content']

    del dataset
    gc.collect()
    
    # text + code
    dataset = load_dataset('m-a-p/CodeFeedback-Filtered-Instruction', split='train')

    for row in dataset:
        yield row['query'] + '\n' + row['answer']

    del dataset
    gc.collect()

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

    # math
    dataset = load_dataset('ajibawa-2023/Maths-College', split='train')
    
    for row in dataset:
        yield row['instruction'] + '\n' + row['output']
    
    del dataset
    gc.collect()

    # math
    dataset = load_dataset('microsoft/orca-math-word-problems-200k', split='train')

    for row in dataset:
        yield row['question'] + '\n' + row['answer']

    del dataset
    gc.collect()

    # emoji
    dataset = load_dataset('badrex/llm-emoji-dataset', split='train')
    
    for row in dataset:
        yield f'{row["character"]}\n{row["unicode"]}\n{row["short description"]}\n{row["tags"]}\n{row["LLM description"]}'
    
    del dataset
    gc.collect()


bpe = BPE(unk_token=None, fuse_unk=False, byte_fallback=False, ignore_merges=True)
tokenizer = Tokenizer(bpe)

special_tokens = [
    '<unk>',
    '<s>',
    '</s>',
    '<|im_start|>',
    '<|im_end|>',
    'system',
    'user',
    'assistant',
    'tool',

    '<tools>',
    '</tools>',
    '<tool_call>',
    '</tool_call>',
    '<tool_response>',
    '</tool_response>',

    '"arguments"',
    '"name"',

    '<arguments>',
    '</arguments>',
    '<argument>',
    '</argument>',
    '<argument-name>',
    '</argument-name>',
    '<argument-type>',
    '</argument-type>',
    '<argument-value>',
    '</argument-value>',
    '<parameter>',
    '</parameter>',
    '<parameter-name>',
    '</parameter-name>',
    '<parameter-type>',
    '</parameter-type>',
    '<parameter-value>',
    '</parameter-value>',
    '<field>',
    '</field>',
    '<field-name>',
    '</field-name>',
    '<field-type>',
    '</field-type>',
    '<field-value>',
    '</field-value>',
    '<name>',
    '</name>',
    '<type>',
    '</type>',
    '<value>',
    '</value>',
    '<function>',
    '</function>',
    '<function-name>',
    '</function-name>',
    '<function-type>',
    '</function-type>',
    '<function-value>',
    '</function-value>',
]

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

for i in range(64 - len(special_tokens)):
    special_tokens.append(f'<|reserved_{i}|>')

# emoji
dataset = load_dataset('badrex/llm-emoji-dataset', split='train')
emoji_chars = [row['character'] for row in dataset if len(row['character']) == 1]
del dataset

# programming languages
dataset = load_dataset('Tanvir1337/programming-languages', split='train')
programming_languages = [n for row in dataset for n in row['text']]
del dataset

# programming languages keywords
dataset = load_dataset('bigcode/programming-languages-keywords', split='train')
code_keywords = [n for row in dataset for n in row['keywords']]
del dataset

tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False, trim_offsets=True, use_regex=True)

tokenizer.post_processor = 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.ByteLevel(add_prefix_space=False, trim_offsets=True, use_regex=True)

trainer = BpeTrainer(
    vocab_size=38400, # 32768 chars + 5034 emojis
    min_frequency=2,
    special_tokens=special_tokens,
    initial_alphabet=emoji_chars + programming_languages + code_keywords,
)

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

CHATML_CHAT_TEMPLATE = (
    "{% for message in messages %}"
        "{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}"
    "{% endfor %}"
    "{% if add_generation_prompt %}"
        "{{ '<|im_start|>assistant\n' }}"
    "{% endif %}"
)

fast_tokenizer = PreTrainedTokenizerFast(
    tokenizer_object=tokenizer,
    chat_template=CHATML_CHAT_TEMPLATE,
    bos_token='<s>',
    eos_token='</s>',
    unk_token='<unk>',
    pad_token='</s>',
    clean_up_tokenization_spaces=False,
)

fast_tokenizer.save_pretrained('../')