license: mit
language:
- en
pretty_name: tokenizer-detect
size_categories:
- n<1K
tags:
- tokenizers
- fertility
- multilingual
- machine-generated-text-detection
configs:
- config_name: fertility
data_files:
- split: train
path: fertility.tsv
tokenizer-detect
Author: Dmitriy Semenkevich (dimhold)
This dataset holds tokenizer fertility (tokens per whitespace word) over FLORES-200 devtest for 204 languages and 4 tokenizers (cl100k_base, o200k_base, gpt2, xlm-roberta-base), the X axis of a test of whether fertility explains machine text detector quality.
Headline, from the repository: the answer to that question is "No": "the sign of the correlation is set by the language the detector was trained on", with partial rho -0.884 against fertility for MULTITuDE detectors fine tuned on English and +0.747 for those fine tuned on Russian.
This is a mirror. The canonical home is the Zenodo record and the GitHub repository; criteria, limitations and full tables live there.
- Code and full write up: https://github.com/dimhold/tokenizer-detect
- DOI: https://doi.org/10.5281/zenodo.22417816
Files
| path | what it is |
|---|---|
fertility.tsv |
816 rows, 204 languages x 4 tokenizers: sentences, words, characters, tokens, tokens per word, tokens per character, single character token share |
How it was produced
Counted by fertility.py (in the repository) over the 1012 FLORES-200 devtest sentences; this file holds counts only, no sentence text. Not mirrored here: the detector quality tables copied from MULTITuDE and Pangram 4, the CommonCrawl language shares and the join outputs, since they reproduce third party published tables. Get them from the GitHub repository. The scripts that turn these files into every number are in the GitHub repository.
Licence
MIT, the licence of the repository, covering code and data.
Citation
@misc{semenkevich2026tokenizerdetect,
author = {Semenkevich, Dmitriy},
title = {tokenizer-detect: tokenizer fertility over 204 languages},
year = {2026},
doi = {10.5281/zenodo.22417816},
url = {https://doi.org/10.5281/zenodo.22417816}
}