Released as a preprint on arXiv.
The paper builds Multi-SFT, a commercially permissive collection of 502,818 instruction-response pairs in five European languages, made through translation and regeneration pipelines. Fine-tuning nine models (1B to 32B) from the EuroLLM, OLMo 2 and Qwen3 families and testing them on 13 benchmarks shows that Multi-SFT alone is competitive with larger data mixes and official instruction-tuned releases, with clear gains on safety tasks.
Authors: Daniela Baiamonte, Kamyar Zeinalipour, Elena Fano, Stefano Simonazzi, Andrea Zugarini.
Read the paper | Models and data
All the lab's language releases are listed in Language models and datasets.