Multi-FLAN-P3

Overview

This dataset is a multilingual subset of the P3 component of the original FLAN collection.

The original FLAN collection is extremely large and aggregates many instruction-following datasets across tasks and domains (see the original FLAN v2 repo for more information). In contrast, this release contains a selected subset of the original FLAN data. The selection strategy mirrors the filtering and sampling used in the flan_v2_converted dataset.

The purpose of this dataset is to provide a manageable, high-quality, and reusable multilingual instruction-following dataset.

All selected entries were translated into multiple languages using permissively licensed large language models, enabling reuse of the generated data—including for training other models.

Languages and Size

Language	Subset	Percentage	Number of Examples
English	en	40%	2,451
Italian	it	15%	883
French	fr	15%	904
Spanish	es	15%	854
German	de	15%	908

English represents 40% of the dataset; the remaining 60% is evenly split across the other four languages.

Translation and Licensing

All non-English entries were generated by translating selected English entries using permissively licensed LLMs:

• Mistral-Small-3.1-24B-Instruct-2503
• Qwen3-VL-235B-A22B-Instruct

The use of permissively licensed models ensures that all translated outputs can be freely reused, including for training other language models, in both research and commercial contexts.

Intended Use

This dataset is intended for:

• multilingual instruction-following training,
• cross-lingual generalization research,
• benchmarking instruction-tuned models,
• downstream model training using permissively licensed data.

Licensing

This dataset is released under ODC-BY. Translated outputs were produced using permissively licensed models and may be reused accordingly.