A Byte-Level BPE tokenizer trained on ['arb_Arab', 'ces_Latn', 'cmn_Hani', 'dan_Latn', 'deu_Latn', 'ell_Grek', 'fra_Latn', 'fw_edu', 'hun_Latn', 'ind_Latn', 'ita_Latn', 'jpn_Jpan', 'nld_Latn', 'pol_Latn', 'por_Latn', 'rus_Cyrl', 'spa_Latn', 'swe_Latn', 'tur_Latn', 'vie_Latn'] data from Fineweb-2-HQ.

Training Details

Parameter	Value
Algorithm	Byte-Level BPE
Language	`['arb_Arab', 'ces_Latn', 'cmn_Hani', 'dan_Latn', 'deu_Latn', 'ell_Grek', 'fra_Latn', 'fw_edu', 'hun_Latn', 'ind_Latn', 'ita_Latn', 'jpn_Jpan', 'nld_Latn', 'pol_Latn', 'por_Latn', 'rus_Cyrl', 'spa_Latn', 'swe_Latn', 'tur_Latn', 'vie_Latn']`
Target Vocab Size	128,000
Final Vocab Size	128,726
Pre-tokenizer	gpt4
Number handling	ltr_3digit
Contraction handling	False
Normalizer	NFC
Special Tokens	`<s>`, `</s>`, `<pad>`, `<unk>`
Training Shards	40, ['fineweb_2_hq.arb_Arab.chunk.00.jsonl', 'fineweb_2_hq.arb_Arab.chunk.01.jsonl', 'fineweb_2_hq.ces_Latn.chunk.00.jsonl', 'fineweb_2_hq.ces_Latn.chunk.01.jsonl', 'fineweb_2_hq.cmn_Hani.chunk.00.jsonl', 'fineweb_2_hq.cmn_Hani.chunk.01.jsonl', 'fineweb_2_hq.dan_Latn.chunk.00.jsonl', 'fineweb_2_hq.dan_Latn.chunk.01.jsonl', 'fineweb_2_hq.deu_Latn.chunk.00.jsonl', 'fineweb_2_hq.deu_Latn.chunk.01.jsonl', 'fineweb_2_hq.ell_Grek.chunk.00.jsonl', 'fineweb_2_hq.ell_Grek.chunk.01.jsonl', 'fineweb_2_hq.fra_Latn.chunk.00.jsonl', 'fineweb_2_hq.fra_Latn.chunk.01.jsonl', 'fineweb_edu_100bt.chunk.00.jsonl', 'fineweb_edu_100bt.chunk.01.jsonl', 'fineweb_2_hq.hun_Latn.chunk.00.jsonl', 'fineweb_2_hq.hun_Latn.chunk.01.jsonl', 'fineweb_2_hq.ind_Latn.chunk.00.jsonl', 'fineweb_2_hq.ind_Latn.chunk.01.jsonl', 'fineweb_2_hq.ita_Latn.chunk.00.jsonl', 'fineweb_2_hq.ita_Latn.chunk.01.jsonl', 'fineweb_2_hq.jpn_Jpan.chunk.00.jsonl', 'fineweb_2_hq.jpn_Jpan.chunk.01.jsonl', 'fineweb_2_hq.nld_Latn.chunk.00.jsonl', 'fineweb_2_hq.nld_Latn.chunk.01.jsonl', 'fineweb_2_hq.pol_Latn.chunk.00.jsonl', 'fineweb_2_hq.pol_Latn.chunk.01.jsonl', 'fineweb_2_hq.por_Latn.chunk.00.jsonl', 'fineweb_2_hq.por_Latn.chunk.01.jsonl', 'fineweb_2_hq.rus_Cyrl.chunk.00.jsonl', 'fineweb_2_hq.rus_Cyrl.chunk.01.jsonl', 'fineweb_2_hq.spa_Latn.chunk.00.jsonl', 'fineweb_2_hq.spa_Latn.chunk.01.jsonl', 'fineweb_2_hq.swe_Latn.chunk.00.jsonl', 'fineweb_2_hq.swe_Latn.chunk.01.jsonl', 'fineweb_2_hq.tur_Latn.chunk.00.jsonl', 'fineweb_2_hq.tur_Latn.chunk.01.jsonl', 'fineweb_2_hq.vie_Latn.chunk.00.jsonl', 'fineweb_2_hq.vie_Latn.chunk.01.jsonl']

Usage

python

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("flexitok/bpe_gpt4_AllL_128000")
tokens = tokenizer.encode("Hello, world!")

Files

tokenizer.json — Full HuggingFace tokenizer
vocab.json — Vocabulary mapping
merges.txt — BPE merge rules

Sample Encoding

Text	Tokens	Token IDs
`Hello, world! 12345 This is a test. こんにちは`	`H, ello, ,, Ġworld, !, Ġ, 123, 45, ĠThis, Ġis, Ġa, Ġtest, ., Ġ, ãģĵ, ãĤĵãģ«, ãģ¡ãģ¯`	`42, 4241, 14, 7657, 3, 223, 29624, 5148, 4945, 592, 266, 2805, 16, 223, 1096, 105401, 126389`

Command used to create this tokenizer:

['/home/gsa/tokenizers2/flexitok/tokenizer_training/train_tokenizers.py', 'algorithm=bpe', 'vocab_size=128000', 'langs=[arb_Arab,ces_Latn,cmn_Hani,dan_Latn,deu_Latn,ell_Grek,fra_Latn,fw_edu,hun_Latn,ind_Latn,ita_Latn,jpn_Jpan,nld_Latn,pol_Latn,por_Latn,rus_Cyrl,spa_Latn,swe_Latn,tur_Latn,vie_Latn]', 'data_dir=/scratch/gsa/data/flexitok/', 'output_dir=/scratch/gsa/trained_tokenizers', 'pretokenizer=gpt4', 'number_handling=ltr_3digit', 'add_numbers=true', 'handle_contractions=false', 'unicode_normalization=nfc', 'use_byte_level_regex=false', 'byte_fallback=false', 'strip_zero_width=false', 'cjk_char_split=false', 'cjk_char_coverage=0.999', 'add_cjk_chars=true', 'max_lines=500_000', 'hf.publish_to_hf=true', 'hf_repo_prefix=flexitok/', 'hf.hf_repo_id=flexitok/bpe_gpt4_AllL_128000', 'hf.collections=[flexitok/script-based-tokenizers]']