Reproduce “Building a Strong Instruction Language Model for a Less-Resourced Language” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Domen Vreš, Tjaša Arčon, Timotej Petrič, Dario Vajda, Marko Robnik-Šikonja, Iztok Lebar Bajec

GaMS3-12B adapts Gemma 3 12B for Slovene through continued pretraining on Slovene and related-language text, parallel English–Slovene data, and instruction tuning. Across ten benchmarks, average rank is 4.25 versus 6.30 for Gemma 3 12B, with lower ranks better; it wins on almost all tasks, but not GSM8K or TruthfulQA. A dated arena snapshot shows 61.2% wins from 95 votes, not superiority.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.