tokentax — The Tokenizer Cost Penalty Across Languages

Class: Dataset + measurement tool Citation anchor: arXiv/ACL paper + versioned dataset DOI

tokentax quantifies the "token tax": the same sentence costs 1× tokens in English but often 2–8× in Tamil, Amharic, or Burmese, because LLM tokenizers are trained on English-heavy corpora. It computes tokenization premium, effective context-window shrinkage, and API cost multipliers for major tokenizers across 100+ languages on parallel corpora, with confidence intervals via evalci.

>>> import tokentax
>>> tok = tokentax.load_tokenizer("gpt-4o")
>>> corpus = tokentax.corpora.load_toy_corpus()
>>> tokentax.premium_report(tok, corpus["tam_Taml"], corpus["eng_Latn"], language="tam_Taml")
PremiumReport(tokenizer='gpt-4o', language='tam_Taml', premium=1.964, 95% CI=[1.673, 2.256], n=4)

Status

Phase 1 (core pipeline) and Phase 2 (full FLORES-200 sweep) are done. The pipeline is validated: it reproduces Petrov et al. 2023's published premium ratios (GPT-2/GPT-4 tokenizers, five languages) within 1.1% — see scripts/validate_against_petrov2023.py. The full sweep covers 8 tokenizers (gpt-4o, gpt-4, gpt-2, llama-3, qwen2.5, deepseek-v3, mistral, gemma-2 — Claude excluded, see below) × 203 non-English FLORES-200 languages, 1.64M raw per-sentence rows, summarized with evalci-backed CIs in data/results/tokentax_summary_devtest.csv. Newer tokenizers show a markedly lower mean premium (gpt-4o/gemma-2 ≈ 2.1×) than legacy ones (gpt-2 ≈ 4.5×) — the "are newer tokenizers fairer" trend the paper will explore.

The OPUS domain-robustness check (scripts/run_opus_domain_check.py) is also done, for the top 30 languages by mean FLORES premium: 21/30 have a mapped second-domain corpus (Helsinki-NLP/opus-100 and/or davidstap/biblenlp-corpus-mmteb, religious register); the other 9 (Shan, Santali, Dzongkha, Tamasheq, Central Atlas Tamazight, Lao, Tigrinya, Manipuri, Kabiyè) have no modern, ungated, non-loading-script parallel corpus available for a second domain at all — that gap is itself a finding (the languages with the highest token tax also have the least data to cross-validate it). Across 264 (language, domain, tokenizer) comparisons, median relative difference from the FLORES estimate is 14% — the premium mostly replicates across register, with some corpus-specific outliers worth a caveat in the paper (Kannada on OPUS-100 diverges up to 106% for gpt-4o; Uyghur and Sanskrit's Bible-corpus pairings diverge 44–72%, likely small/idiosyncratic samples in those specific files). Results in data/results/opus_domain_check.csv.

Phase 3 (results dataset + explorer) is published: shreyaskc/tokentax-results-v1 on Hugging Face — flores200_summary.parquet (the CI-backed long-format table), flores200_raw.parquet (1.64M per-sentence rows), and opus_domain_check.parquet, under CC0 for the derived statistics (no source sentence text is redistributed). Re-publish with scripts/publish_hf_dataset.py.

The tokentax explorer (heatmap + "what does N tokens buy in your language" calculator) is live as a static Space — Hugging Face requires a PRO subscription to host Gradio/Docker Spaces even on the free CPU tier, so hf_space/ is a Plotly.js + vanilla-JS page reading a bundled data.json/pricing.json snapshot rather than a live Hub read. Rebuild with scripts/build_static_explorer.py after a new results version, then redeploy with scripts/publish_hf_space.py. A functionally identical Gradio version (live Hub reads, no rebuild step) sits unpublished at scripts/gradio_explorer_prototype/ for if the account upgrades to PRO.

Released on PyPI: tokentax 0.1.0.

Paper drafted: paper/main.tex (8 pages) reproduces Petrov et al. 2023 within 1.1%, reports the full 8-tokenizer sweep, the domain-robustness check, and a cost-analysis table, with an Ethics Statement and Limitations section. Not yet submitted to arXiv — that step needs an arXiv account and is a manual submission, not something scripted. Regenerate the figures with paper/scripts/make_figures.py and compile with pdflatex → bibtex main → pdflatex × 2 from within paper/.

Not yet released: Zenodo DOI. See PLANNING.md for the full implementation brief and remaining release checklist (Papers with Code, seed emails, workshop poster).

Two registry notes from the real run: llama-3 resolves to the NousResearch/Meta-Llama-3-8B mirror, not meta-llama/Meta-Llama-3-8B, because the official repo requires Meta's manual license approval rather than an instant click-through; mistral turned out not to be gated at all (needs protobuf installed, not a license). Claude is excluded from the sweep by choice, not necessity — see registry.py if you want to re-add it with ANTHROPIC_API_KEY set.

Install

pip install tokentax

To hack on the library itself (and run the test suite):

git clone https://github.com/Shreyaskc/token-tax.git
cd token-tax
pip install -e ".[test,corpora]"
pytest tests/

Requires Python ≥3.9. Core runtime deps: numpy, pandas, pyyaml, evalci, tiktoken, transformers, sentencepiece, huggingface_hub. Optional extras: corpora (Hugging Face datasets, for pulling FLORES-200) and claude (the anthropic client, for Claude's token-count API).

One-time setup for gated resources

Several tokenizers (Llama 3, Mistral, Gemma) and the FLORES-200 corpus (facebook/flores) are gated on Hugging Face — a free, auto-approved license click, not a manual review:

1. Visit the dataset/model page while logged in and click "Agree and access repository".
2. huggingface-cli login (or set HF_TOKEN).

Claude has no downloadable tokenizer; tokentax uses Anthropic's free token-count API instead, which needs ANTHROPIC_API_KEY.

Usage

List registered tokenizers

tokentax list-tokenizers

Premium for one language vs. English

tokentax premium gpt-4o tam_Taml --corpus toy   # bundled toy corpus, no network
tokentax premium gpt-4o tam_Taml --corpus flores200   # full FLORES-200 (needs the gate above)

Library API

import tokentax

tok = tokentax.load_tokenizer("gpt-4o")

# per-sentence metrics
tokentax.metrics.premium_ratios(tok, tamil_sentences, english_sentences)
tokentax.metrics.bytes_per_token(tok, sentences)
tokentax.metrics.chars_per_token(tok, sentences)
tokentax.metrics.effective_context_window(context_tokens=128_000, tokens_per_word=2.4)

# CI-backed report for one (tokenizer, language) pair, or a whole corpus
tokentax.premium_report(tok, tamil_sentences, english_sentences, language="tam_Taml")
tokentax.premium_table(tok, tokentax.corpora.load_toy_corpus())

# $/token pricing (placeholder figures — see pricing.yaml)
tokentax.pricing.estimate_cost("gpt-4o", n_tokens=1_000_000)

What's validated, and how

• Confidence intervals are computed by evalci.ci(method="bootstrap"), not reimplemented — the same statistically-validated routine used across this portfolio's benchmarks.
• Premium ratios are computed on FLORES-200's aligned parallel sentences (same meaning across languages), not independent monolingual corpora — the only methodologically defensible basis for a cross-language ratio.
• pricing.yaml ships with verified: false; tokentax.pricing.load_pricing() warns until it's checked against live provider pricing and flipped to true. Do not cite a dollar figure from an unverified snapshot.

Release checklist

CITATION.cff (done) → full FLORES-200 run + validation against Petrov et al. 2023 (done) → HF dataset + Space (done) → PyPI (done) → arXiv (drafted, submission pending) → peer-reviewed submission (ACL Rolling Review → Findings of ACL/EMNLP/NAACL, COLING, or a fitting workshop — MRL, LoResMT, VarDial, ComputEL; arXiv alone does not satisfy EB-1A's "scholarly articles" criterion, peer review does) → Papers with Code → Zenodo → seed emails to multilingual-NLP and AI-policy researchers → workshop poster.