Finding

Sign in with GitHub
← Publications

Finding · P53 · Author-curated

APT4 needs more tokens per word than APT3 on all seven corpora and on both preamble texts.

Published by @stw2 · 2026-09-14 · Sources, measurements and interpretation are supplied by the author.

Structured assertion

Relation: Higher on a count of texts

metric
Fertility ratioQuantity compared.
subject
APT4 tokenizerTokenizer compared.
comparator
APT3 tokenizerTokenizer compared with.
count higher
9 textsTexts where the subject's value is higher.
count total
9 textsTexts compared.
texts
seven corpora; Polish preamble; English preambleTexts compared.

Experimental provenance

Method and evaluation protocol
Compare APT4 and APT3 tokens per word on each corpus and on the preamble protocol texts.
Dataset
Seven corpora and the preamble protocol texts.Version: unspecified · Access: restricted
Reported results
en-arxiv-abstracts: 2.462 against 2.439; en-latex-methods: 3.976 against 3.895; en-python-code: 6.626 against 6.329; en-gsm8k: 1.896 against 1.820; pl-science-pes: 1.957 against 1.778; pl-wiki-science: 2.026 against 1.856; pl-informal: 1.528 against 1.497; preamble-pl: 1.616 against 1.483; preamble-en: 1.984 against 1.943
Uncertainty and replication
Point values; per-corpus intervals are in results/atlas.json.
Limitations
APT4's training recipe is unpublished, so the cause is not identified.

Paper citations

arXiv:2604.10799v1 →Revision supplied by author

Associated through an explicitly referenced cited claim.

Concept definitions

Reuse the defining version and key when the meaning fits your assertion.

Higher on a count of texts

The subject's metric is higher than the comparator's on count_higher of count_total texts.

Key higher_on_count · version d12932ff-15a4-4030-ba68-04695fc77ae9

APT3 tokenizer

Polish tokenizer introduced with the Polish APT3 model.

Key apt3 · version a59f02ac-4303-44b5-a5b0-5decab214308

Fertility ratio

Average number of tokens required to represent a text.

Key fertility_ratio · version f3b88c39-b184-4cb7-85e6-916f260ad0ad

APT4 tokenizer

Polish-optimised tokenizer of the Bielik v3 PL models.

Key apt4 · version d497f94d-5373-4652-887c-55c001b6472c

Exact references

related

e150c2d5-b823-43d9-9440-5f6428448be1

APT4 extends APT3's design; its fertility is higher on every text measured.