gemma-4-31B-it-Adetayo
Icelandic fine-tune of Google Gemma 4 31B instruction-tuned, targeting Icelandic morphology and grammar.
Scores (Miðeind Icelandic LLM leaderboard, official run)
| score | |
|---|---|
| this model, 6-task average | 70.09 |
| base gemma-4-31b-it, 6-task average | 71.24 |
| this model, 5-task local average (thinking off) | 83.59 |
Read those two top numbers carefully. The base is a reasoning model and its leaderboard run uses its thinking path; this fine-tune is submitted and evaluated with thinking disabled. On the 5-task subset, thinking-on is worth roughly 7 points on this base (85.0 with, 77.92 without), so the fine-tune and the base entry are not measured the same way. Where they are comparable, morphology is the gain.
A reasoning-preserving variant was also trained on rejection-sampled inflection traces. It held Wino, GED, Belebele and ARC exactly but moved inflection not at all, because 293 of 354 traces covered cases the base already solved. The method is sound, the data was too easy, so that variant was not published. This model is at its supervised fine-tuning ceiling.
Use
Standard text generation. Apply the Gemma chat template (tokenizer.apply_chat_template). Load the text path on GPU (Gemma4ForConditionalGeneration, .to("cuda")). Evaluate with thinking disabled.
License
Gemma derivative. Use is governed by the Gemma Terms of Use and the Gemma Prohibited Use Policy. "Gemma" is retained in the model name as required.
Training data and methodology are proprietary and are not distributed with the model.