Preregistration: Token cost of Norwegian Bokmål and Nynorsk relative to English across LLM tokenizers (confirmatory study)
Preregistration of a confirmatory study of how many more tokens Norwegian Bokmål and Nynorsk text costs than English across LLM tokenizers (Anthropic Claude, OpenAI o200k_base and cl100k_base, Gemma 4 / Borealis 2, NorMistral-11B). Ten hypotheses on the parity premium (Petrov et al. 2023) are decided on FLORES-200 dev with a paired bootstrap and 99.5 % Bonferroni intervals; UD r2.19 test sets and Målfrid parallel text are reported descriptively. An earlier exploratory pilot on FLORES-200 devtest (tag pilot-v1) and the locked analysis code (signed tag prereg-v1) are included as git archives with OpenTimestamps proofs.
Authors
- Eirik Botten Nicolaysen (ORCID: https://orcid.org/0009-0001-9188-6788)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24189351
- Citations
- 1
- Primary Topic
- Natural Language Processing Techniques
- Type
- article
- Field-Weighted Citation Impact
- 1.39