Model components / Training & synthetic datasets

FLAN Collection (v2)

Google Research's instruction-tuning compilation of more than 1,800 tasks, drawing on P3, Super-NaturalInstructions and chain-of-thought data, which trained Flan-T5 and Flan-PaLM. It remains a building block for later instruction mixes, folded into Tulu 3, OLMo 3's Dolci and Marin's Dolmino.

The recipe repository is the primary artifact; the widely used Hugging Face conversion carries a card documenting the conversion rather than the tasks. Verified 2026-08-13 via the google-research/FLAN repository and the ai2-adapt-dev/flan_v2_converted dataset card.

Openness

3 medium confidence

3.0

access: ungated(conversion)
license: Apache-2.0(recipe)+per-task
dataset_card: partial

Released as an Apache-2.0 recipe and repository, and the widely used Hugging Face conversion (ai2-adapt-dev) is ungated. The underlying task datasets keep their own licenses, so the Apache-2.0 covers the recipe rather than the corpus and the license defers to its components, as the-pile and stack-edu do. A deferred license plus a partial dataset card puts this at 3.

https://github.com/google-research/FLAN recorded 2026-08-13
Apache-2.0 FLAN v2 recipe
https://huggingface.co/datasets/ai2-adapt-dev/flan_v2_converted recorded 2026-08-13
ungated HF conversion used by Tulu/OLMo mixes

Adoption

1 high confidence

1.0

108 downloads in the trailing 30 days for ai2-adapt-dev/flan_v2_converted, which bands at <1K, level 1 on the dataset adoption scale.

https://huggingface.co/api/datasets/ai2-adapt-dev/flan_v2_converted recorded 2026-08-13
108 downloads in the trailing 30 days for ai2-adapt-dev/flan_v2_converted

Capability

3 high confidence

3.0

Strong historical comparative ablations (Flan 2022 beats P3/T0/SNI on the same T5-XL) and still a component in Tulu 3, Dolci, and Marin's mixes, but dated as a standalone chat-era recipe. The Flan Collection reports Flan-T5 outperforming prior work by 3-17%+ across evaluation settings, and the Dolci SFT card still consumes FLAN v2.

https://arxiv.org/abs/2301.13688 recorded 2026-08-13
Abstract - ablations on the Flan Collection show Flan-T5 outperforming prior work by 3-17%+ across evaluation settings. The per-collection P3/Super-NaturalInstructions rows are in the paper body, not on this page
https://huggingface.co/datasets/allenai/Dolci-Instruct-SFT recorded 2026-08-13
OLMo 3 Dolci still consumes FLAN v2 (89,981 prompts)

Verified 2026-08-13