Multi-SFT builds permissive instruction data across five languages

Can one instruction dataset support training across languages while avoiding proprietary-generated responses whose terms can restrict competing training? Many existing resources fall short on English-heavy coverage or licensing and provenance.

Multi-SFT brings together multilingual examples selected and produced to address both constraints, then tests how models trained on it perform.

Multi-SFT combines expert-checked translation and regeneration to build a permissively licensed instruction dataset across five European languages.

Researchers trained nine models from EuroLLM, OLMo 2, and Qwen3 on Multi-SFT, then assessed them on multilingual tests of tasks including translation, reading comprehension, and question answering.

They averaged results across languages within each test and combined the test scores into a multilingual average.

Multi-SFT-trained models had a higher average than corresponding official instruction-tuned models in six of nine comparisons. This supports the collection’s use for these tested models and tasks, though the comparison does not identify which training differences account for the result.

The official models’ starting checkpoints, training data, and objectives are unspecified, so that comparison cannot isolate Multi-SFT’s effect; Qwen3 was tested without thinking mode. Human checks covered samples, and transfer beyond the five included European languages remains unknown.