The challenge
Kinyarwanda text needs more tokens than its English source. With Gemma 3, the same 1,012 FLORES+ sentences take 53,301 tokens, against 27,114 in English. About 12.1 million people speak it as a first language (Ethnologue, 2019).
The change
Nzinga added 32,000 African word pieces to Gemma 3's tokenizer, trained on open web text in 29 languages, including 25.0 million characters of Kinyarwanda. Every Gemma 3 token kept its ID, and no test sentence was in training.
flores+:
sentences: 1012
nzinga: 42,162
gpt-4o: 46,011
overlap_13gram: 0
Closest to the target, with Somali, Igbo and Oromo.
Tier 1 results
,
FLORES+ devtest premium
,
Kinyarwanda, 1.56x
The result
Nzinga needs 8.4% fewer tokens than GPT-4o on FLORES+ and 10.0% fewer on NTREX-128. The premium is 1.56 times English, just above the target.
