The follwing information is gathered by Chris Ha for research purposes. No guarantee of correctness implied.
For specific information, consult original developers or refer to the Original license and original_README.md.
Architecture
- likely sentencepiece trained tokenizer transfered imported into huggingface tokenizers. If so, original sentencepiece model not provided.
Algorithm
- BPE
- UOV handled with byte_fallback infered by the absence of bytelevel() and spm origin
Pretokenizers
- Whitespace : (Can be relevant for code or compression.)
- Metaspace() with ("▁" U+2581 LOWER ONE EIGHTH BLOCK) as meta symbol
- prefix applied (i.e. both "the" and "▁the" present)
- Whitespace reservation : None (no multiple whitespace nor metasymbols)
- Punctuation : (code & math)
- Split and Isolated
- Numbers : (code & math, compression ratios)
- Split and Isolated(invidiual_digits=true)
Vocabulary
- Size : 100257
- Around 96567 are single width character unicode characters (WIP)
- Average Length : WIP
- Median Length : 1 char
- Languages
- "multiple languages including ar, bg, ca, cs, da, de, el, en, es, et, fa, fi, fr, hi, hu, id, it, iw, ja, kk, ko, lt, lv, mr, ms, nl, no, pl, pt, ro, ru, sk, sl, sr, sv, ta, th, tr, uk, vi, and zh"
- Most multi-char tokens are english or chinese