import gc import sys # import string 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(): # 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 dataset = load_dataset('nampdn-ai/tiny-textbooks', split='train') for row in dataset: yield row['text'] del dataset gc.collect() ## text # dataset = ( # load_dataset('wikimedia/wikisource', lang, split='train') # for lang in ['20231201.ar', '20231201.as', '20231201.az', '20231201.ban', '20231201.be', '20231201.bg', '20231201.bn', '20231201.br', '20231201.bs', '20231201.ca', '20231201.cs', '20231201.cy', '20231201.da', '20231201.de', '20231201.el', '20231201.en', '20231201.eo', '20231201.es', '20231201.et', '20231201.eu', '20231201.fa', '20231201.fi', '20231201.fo', '20231201.fr', '20231201.gl', '20231201.gu', '20231201.he', '20231201.hi', '20231201.hr', '20231201.hu', '20231201.hy', '20231201.id', '20231201.is', '20231201.it', '20231201.ja', '20231201.jv', '20231201.kn', '20231201.ko', '20231201.la', '20231201.li', '20231201.lij', '20231201.lt', '20231201.mk', '20231201.ml', '20231201.mr', '20231201.nap', '20231201.nl', '20231201.no', '20231201.or', '20231201.pa', '20231201.pl', '20231201.pms', '20231201.pt', '20231201.ro', '20231201.ru', '20231201.sa', '20231201.sah', '20231201.sk', '20231201.sl', '20231201.sr', '20231201.su', '20231201.sv', '20231201.ta', '20231201.te', '20231201.th', '20231201.tr', '20231201.uk', '20231201.vec', '20231201.vi', '20231201.wa', '20231201.yi', '20231201.zh', '20231201.zh-min-nan'] # ) # # for d in dataset: # for row in d['text']: # yield row # # 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() ## text # dataset = ( # load_dataset('csebuetnlp/xlsum', lang, split='train') # for lang in ['amharic', 'arabic', 'azerbaijani', 'bengali', 'burmese', 'chinese_simplified', 'chinese_traditional', 'english', 'french', 'gujarati', 'hausa', 'hindi', 'igbo', 'indonesian', 'japanese', 'kirundi', 'korean', 'kyrgyz', 'marathi', 'nepali', 'oromo', 'pashto', 'persian', 'pidgin', 'portuguese', 'punjabi', 'russian', 'scottish_gaelic', 'serbian_cyrillic', 'serbian_latin', 'sinhala', 'somali', 'spanish', 'swahili', 'tamil', 'telugu', 'thai', 'tigrinya', 'turkish', 'ukrainian', 'urdu', 'uzbek', 'vietnamese', 'welsh', 'yoruba'] # ) # # for d in dataset: # for row in d['text']: # yield row # # del dataset # gc.collect() ## text # dataset = load_dataset('recursal/SuperWikiNEXT-32B', split='train') # # for row in dataset['text']: # yield row # # del dataset # gc.collect() # 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() ## code # dataset = load_dataset('nampdn-ai/tiny-codes', split='train') # # for row in dataset: # yield row['prompt'] + '\n' + row['response'] # # 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() # text dataset = load_dataset('mlabonne/FineTome-100k', split='train') for row in dataset['conversations']: yield '\n'.join(n['value'] for n in row) del dataset gc.collect() # instruction dataset = load_dataset('arcee-ai/agent-data', split='train') for row in dataset['conversations']: yield '\n'.join(n['value'] for n in row) del dataset gc.collect() # instruction dataset = ( load_dataset('cognitivecomputations/SystemChat-2.0', data_files='SystemChat_filtered.jsonl', split='train'), load_dataset('cognitivecomputations/SystemChat-2.0', data_files='SystemChat_multilingual.jsonl', split='train'), ) for d in dataset: for row in d['messages']: yield '\n'.join(n['content'] for n in row) 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='', fuse_unk=True, byte_fallback=True) tokenizer = Tokenizer(bpe) special_tokens = [ '', '', '', '<|im_start|>', '<|im_end|>', 'system', 'user', 'assistant', 'tool', '', '', '', '', '', '', '"arguments"', '"name"', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', '', ] for i in range(2, 25): special_tokens.append(' ' * i) for i in range(128 - 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=32768, # 2 ** 15 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='', eos_token='', unk_token='', pad_token='', clean_up_tokenization_spaces=False, ) fast_tokenizer.save_pretrained('../')