Abzalbek89/kk-tokenizer-fertility-baseline

Kazakh Tokenizer Fertility Baseline

Reproducible fertility benchmark of subword tokenizers on the Kazakh language. Companion artifact for the paper "Tokenizer Optimization for Kazakh Small Language Models" (in preparation, target: ACM TALLIP).

Headline numbers

	Tokenizer	Fertility
🥇 Best overall	kk-bpe-32k	1.679
🚨 Worst	GPT-4 (cl100k)	5.895
GPT-4 penalty	GPT-4 (cl100k) is 3.51× worse than the best Kazakh-trained tokenizer

→ The custom Kazakh BPE tokenizer is 3.51× more efficient than GPT-4 for Kazakh text, which translates to ~72% lower API cost and ~3.5× longer effective context window.

Full ranking (13 tokenizers)

Ours — Kazakh-trained, no morphology pre-segmentation

Tokenizer	Vocab	Fertility ↓	Chars/Token ↑
kk-bpe-32k	32,000	1.679	4.672
kk-sp-bpe-32k	32,000	1.758	4.463
kk-sp-unigram-32k	32,000	1.762	4.454
kk-hf-unigram-32k	32,000	2.026	3.872

Ours — Kazakh-trained, morphology-aware (Morfessor → tokenizer)

Tokenizer	Vocab	Fertility ↓	Chars/Token ↑
kk-morph-hf-bpe-32k	32,000	1.761	4.456
kk-morph-sp-unigram-32k	32,000	1.945	4.034
kk-morph-hf-unigram-32k	32,000	2.346	3.345

Reference — multilingual / industry tokenizers

Tokenizer	Vocab	Fertility ↓	Chars/Token ↑
XLM-R	250,002	2.181	3.597
GPT-4o (o200k)	200,019	2.599	3.020
mBERT	119,547	2.929	2.679
Qwen2.5	151,643	4.786	1.639
Llama-3	128,000	4.803	1.634
GPT-4 (cl100k)	100,277	5.895	1.331

Methodology

Held-out evaluation set: 2,966,207 whitespace-words (≈15K documents) from the validation split of Abzalbek89/corpus_clean.

Metrics:

• Fertility = total tokens / total whitespace-words (lower is better).
• Compression (chars/token) = total characters / total tokens (higher is better).
• Compression (bytes/token) = total UTF-8 bytes / total tokens (higher is better).

Trained tokenizers (vocab = 32,000, all share the same training corpus):

• kk-bpe-32k — HF tokenizers ByteLevel BPE
• kk-sp-bpe-32k — SentencePiece BPE (character_coverage=1.0, NFKC)
• kk-sp-unigram-32k — SentencePiece Unigram
• kk-hf-unigram-32k — HF tokenizers Unigram with ByteLevel pre-tok
• kk-morph-hf-bpe-32k — HF BPE on Morfessor-segmented corpus
• kk-morph-hf-unigram-32k — HF Unigram on Morfessor-segmented corpus
• kk-morph-sp-unigram-32k — SentencePiece Unigram on Morfessor-segmented corpus

Reference tokenizers: mBERT, XLM-R, Llama-3, Qwen 2.5, GPT-4 (cl100k_base), GPT-4o (o200k_base).

Files

Path	Description
experiment.py	v1 — single Unigram baseline
experiment_v2.py	v2 — extended (BPE/Unigram across libraries)
experiment_v3_fix.py	v2 fix — re-measure SentencePiece tokenizers via raw SentencePieceProcessor
experiment_morph.py	Experiment 2 — morphology-aware variants
morfessor.bin	trained Morfessor segmentation model (top-1M Kazakh words)
v2/fertility_v2.csv, .json, .png	fixed v2 numbers
v3/fertility_v3.csv, .json, .png	extended v3 numbers (with morph-aware)
v3/RESULTS_V3.md	human-readable v3 report

Reproducing

On a vast.ai / RunPod / Colab instance with ≥30GB disk and ≥16GB RAM:

pip install 'datasets>=2.14' 'transformers>=4.45' 'tokenizers>=0.20' \
            'sentencepiece>=0.2.0' 'huggingface_hub>=0.25' \
            tiktoken matplotlib morfessor

export HF_TOKEN=hf_...   # write access
export HF_USER=Abzalbek89

curl -fsSL -o experiment_morph.py \
  https://huggingface.co/datasets/Abzalbek89/kk-tokenizer-fertility-baseline/resolve/main/experiment_morph.py
python experiment_morph.py

Total runtime: ~30 min on cached Morfessor model, ~70 min cold start.

Companion tokenizer repos

• Abzalbek89/kk-tokenizer-bpe-32k
• Abzalbek89/kk-tokenizer-sp-bpe-32k
• Abzalbek89/kk-tokenizer-sp-unigram-32k
• Abzalbek89/kk-tokenizer-hf-unigram-32k
• Abzalbek89/kk-tokenizer-morph-hf-bpe-32k ¹
• Abzalbek89/kk-tokenizer-morph-hf-unigram-32k ¹
• Abzalbek89/kk-tokenizer-morph-sp-unigram-32k ¹

¹ Morphology-aware variants — apply Morfessor (morfessor.bin in this repo) before encoding.

Citation

@misc{kk_tokenizer_fertility_2026,
  title        = {Kazakh Tokenizer Fertility Baseline},
  author       = {Abzalbek Ulasbek},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Abzalbek89/kk-tokenizer-fertility-baseline}},
}

License

Apache 2.0

---

Last updated: fertility_v3.csv (13 tokenizers)