stukenov's picture
add dataset card
38862c1 verified
|
Raw
History Blame Contribute Delete
1.5 kB
metadata
language:
  - kk
  - ru
  - en
license: mit
library_name: tokenizers
tags:
  - kazakh
  - tokenizer
  - bpe
  - morphology
  - til
pretty_name: Til MorphoBPE 96K (kk/ru/en/code/math)

Til-vocab-morphbpe-96k-kkrumix-v1

Byte-level BPE tokenizer (96 000 vocab) for the Til Kazakh foundation-model program. Kazakh-priority, with Russian, English, code and math coverage.

Summary

  • Vocab size: 96 000 (byte-level BPE, tokenizers).
  • Priority: Kazakh-first; balanced kk/ru shares + en/code/math.
  • Special tokens: <pad>=0, <bos>=1, <eos>=2.
  • Built on: kaznu, from on-disk collected kk/ru (Til corpus exp053/054) + streamed en/code/math (FineWeb-Edu / code / math) — streaming avoids contention with the collectors.

Fertility (tokens/word, lower = better)

Lang Fertility Note
kk 1.699 ✅ target < 2.0
ru 1.619 marginal (aspirational < 1.6; +0.019)
en 1.355

Raising the Russian vocab share dropped ru fertility 1.718 → 1.619 (kk:ru mix tuning).

Usage

from tokenizers import Tokenizer
tok = Tokenizer.from_file("tokenizer.json")
ids = tok.encode("Сәлеметсіз бе, әлем!").ids

Part of the Til program

Companion corpora (same vocab): Til-corpus-tokenized-multi-morphbpe96k-mvp100b-v1 (kk, 8.4B), Til-corpus-tokenized-mix-morphbpe96k-v0 (ru/en/code/math, 12B), Til-corpus-tokenized-kk-extra-v0 (0.83B), Til-corpus-tokenized-kk-madlad-v0 (0.40B).

License: MIT.