Dataset Viewer
Auto-converted to Parquet Duplicate
rows
list
diacritic_surcharge
dict
n_sentences
int64
n_languages
int64
tokenizers
list
[{"tokens":128377,"chars":111639,"bytes":199902,"tokens_per_sentence":126.85474308300395,"bytes_per_(...TRUNCATED)
{"gpt2":{"marked_tokens":105095,"stripped_tokens":49133,"surcharge":2.138990088128142},"bloom":{"mar(...TRUNCATED)
1,012
204
["gpt2","bloom","qwen2.5","nllb","xlm-r","mbert","aya","phi-3","mistral","falcon","gpt-3.5/4 (cl100k(...TRUNCATED)

owóorí: the tokenizer tax, measured

Token-count premiums for the 204 languages of FLORES-200 under 12 tokenizers, from the identical 1,012 professionally translated sentences. The premium is tokens(language) / tokens(English) on the same content, so it reads directly as a price multiplier for API cost, latency and context shrinkage.

Files

  • premiums.csv: one row per language x tokenizer (204 x 12 = 2448 rows): premium, raw token counts, characters, bytes, bytes per token.
  • measurements.json: everything above plus the Yoruba diacritic-surcharge measurements (marked vs stripped text under every tokenizer).

Headlines

  • Santali pays 12.74x English's rate at the frontier (median of cl100k, o200k, Qwen2.5), the worst of the 204.
  • 61% of languages pay more than 2x at the frontier; 17% pay more than 4x.
  • The tax is a design choice: Amharic pays 5.78x under GPT-4o and 1.29x under NLLB; Yoruba pays 2.17x and 1.48x.
  • Writing Yoruba with its tone and dot-below marks costs 1.52x the stripped text under GPT-4o: a second tax, on orthography itself.

Provenance

Source text: FLORES-200 devtest (NLLB Team 2022, CC-BY-SA 4.0), from Meta's public distribution. Tokenizers: GPT-2, Phi-3, Mistral v0.3, Falcon-7B, cl100k_base, o200k_base, Qwen2.5, NLLB-200, XLM-R, mBERT, Aya-101, BLOOM, all loaded from their public repositories. Llama and Gemma are gated and not included. Measurement code and tests are in the GitHub repository.

Downloads last month
26