tokentax-results-v1

Measurements of the "token tax" — the tokenization premium, i.e. how many more tokens a language costs relative to English for the same content — across 8 LLM tokenizers and 203 FLORES-200 languages, with confidence intervals computed via evalci.

Produced by tokentax (commit 0e5e24d). Fully re-runnable: scripts/run_flores_sweep.py → scripts/summarize_with_evalci.py → scripts/run_opus_domain_check.py in that repo regenerate every file here from scratch.

No original sentence text is redistributed

This dataset contains only token counts, premium ratios, and confidence intervals — not the source sentences themselves. Premium ratios were computed against FLORES-200 (facebook/flores, gated, CC-BY-SA-4.0), OPUS-100 (Helsinki-NLP/opus-100), and a Bible-translation corpus (davidstap/biblenlp-corpus-mmteb); none of that text is copied here. See LICENSE for the license of the derived statistics vs. the source corpora.

Files

flores200_summary.parquet

The primary results table: long-format (tokenizer, tokenizer_version, language, corpus, metric, value), one row per metric per (tokenizer, language) pair, from the full FLORES-200 devtest sweep (8 tokenizers × 203 non-English languages).

column	description
tokenizer	one of gpt-4o, gpt-4, gpt-2, llama-3, qwen2.5, deepseek-v3, mistral, gemma-2 (Claude excluded from this release, see Limitations)
tokenizer_version	pinned Hugging Face revision, or None for tiktoken-backed tokenizers
language	FLORES-200 code, e.g. tam_Taml
corpus	flores200_devtest
metric	premium_ratio_estimate, premium_ratio_ci_lower, premium_ratio_ci_upper, or n_sentences
value	the metric's value

premium_ratio is tokens(language sentence) / tokens(aligned English sentence), mean over 1012 aligned FLORES-200 devtest sentences; CIs are a 95% bootstrap interval via evalci.ci(method="bootstrap").

flores200_raw.parquet

Raw per-sentence premium ratios underlying the summary above — 1.64M rows, one per (tokenizer, language, sentence). Kept so anyone can recompute a different CI method, subset languages, or check for outlier sentences without re-tokenizing FLORES-200 themselves.

column	description
tokenizer	tokenizer name
tokenizer_revision	pinned HF revision
language	FLORES-200 code
corpus	flores200_devtest
sentence_idx	position in the devtest split (0–1011)
premium_ratio	tokens(this sentence) / tokens(aligned English sentence)

opus_domain_check.parquet

Domain-robustness check: for the 30 languages with the highest mean premium in flores200_summary, compares the FLORES-200 estimate against a second, non-encyclopedic corpus (OPUS-100 mixed-domain text and/or a Bible-corpus religious-register text), where one is available.

column	description
language	FLORES-200 code
domain	opus100 or bible
tokenizer	tokenizer name
flores_premium	the FLORES-200 devtest estimate for this (tokenizer, language)
domain_premium	the estimate on this domain's own aligned English pair
domain_ci_lower, domain_ci_upper	95% bootstrap CI on domain_premium
n_sentences	sentence count used for this domain's estimate
rel_diff	|domain_premium - flores_premium| / flores_premium
flag	DIVERGES if rel_diff exceeds 35%, else ok

Coverage and known limitations

• Claude is excluded from this release. Its tokenizer has no downloadable artifact (only Anthropic's token-count API), so including it would require every user of this dataset to hold an API key just to reproduce it. See the source repo's registry.py to re-add it.
• llama-3 resolves to the NousResearch/Meta-Llama-3-8B mirror, not meta-llama/Meta-Llama-3-8B — the official repo requires Meta's manual license approval rather than an instant click-through.
• Domain-robustness coverage is partial by construction. Of the 30 highest-premium languages, 9 (Shan, Santali, Dzongkha, Tamasheq, Central Atlas Tamazight, Lao, Tigrinya, Manipuri, Kabiyè) have no modern, ungated, non-loading-script parallel corpus available on Hugging Face for a second domain at all. That gap is reported explicitly in opus_domain_check (absent rows), not silently smoothed over — the languages with the highest token tax also tend to have the least data to cross-validate it. Where a second domain IS available, most (language, tokenizer, domain) triples replicate the FLORES estimate within ~14% (median relative difference across 264 comparisons); a handful of outliers exceed 35% (notably Kannada on OPUS-100, and Uyghur/Sanskrit on the Bible corpus, likely corpus-specific sampling artifacts, not a general finding).
• Reproduces Petrov et al. 2023 ("Language Model Tokenizers Introduce Unfairness Between Languages," NeurIPS 2023) within 1.1% max relative error on GPT-2/GPT-4 across 5 languages — see the source repo's scripts/validate_against_petrov2023.py.

Citation

An arXiv paper is planned; until it's out, cite the software/dataset directly:

@misc{chandrahas2026tokentax,
  title  = {tokentax: The Tokenizer Cost Penalty Across Languages},
  author = {Chandrahas, Shreyas K},
  year   = {2026},
  url    = {https://github.com/Shreyaskc/token-tax}
}

Versioning

This is v1, built from FLORES-200 devtest with the tokenizers listed above. Re-runs after new tokenizer releases will ship as tokentax-results-v2, etc., each independently versioned per the source repo's release policy.