Supervised fine-tuning (SFT) datasets are essential for training instruction-following models, but they remain overwhelmingly English-centric. According to a new arXiv preprint, their content is frequently generated by proprietary models whose terms of service forbid training competing systems, adding a commercial restriction on top of the language imbalance.

The paper introduces Multi-SFT, a permissively-licensed multilingual instruction-following dataset designed to counter this trend. By offering a freely usable alternative, it aims to let researchers and companies train their own systems without the legal barriers that come with data derived from proprietary models.

The announcement is brief and does not yet specify the dataset's size, language coverage, or construction methodology. Still, the stated goal is clear: provide a broadly accessible resource that moves the field away from English-only, restricted training data.