BLOOMZ

BLOOMZ is the instruction-tuned variant of BLOOM, produced by finetuning the 176B multilingual base model on xP3 — a composite of prompted datasets in 46 languages with English and machine-translated prompts. The key finding (arXiv:2211.01786, Nov 2022) was that finetuning large multilingual models on English tasks with English prompts produces task generalization to non-English languages that appear only in the pretraining corpus — capabilities that are both task- and language-agnostic.

BLOOMZ demonstrated that zero-shot generalization need not be an English-only phenomenon: it achieved strong zero-shot results on tasks in languages the model had never intentionally seen, and it introduced the xP3 multitask finetuning recipe that influenced later instruction-tuning work. Weights and training code are public, though the Responsible AI License (RAIL) restricts certain uses.