Multi-SFT: Permissively-Licensed Multilingual Instruction-Following

In short

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.

Open European LLMs

Cite

@misc{baiamonte2026multisftpermissivelylicensedmultilingualinstructionfollowing,
      title={Multi-SFT: Permissively-Licensed Multilingual Instruction-Following}, 
      author={Daniela Baiamonte and Kamyar Zeinalipour and Elena Fano and Stefano Simonazzi and Andrea Zugarini},
      year={2026},
      eprint={2610.22086},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.22086}, 
}

Copy BibTeX

All thirty-one papers