FLAN Collection (v2)
Google ResearchGoogle Research's instruction-tuning compilation 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, including the Tulu 3 SFT mixture and Olmo 3's Dolci.
The recipe repository is the primary artifact; the widely used Hugging Face conversion carries a card documenting the conversion rather than the tasks.
Openness
3 medium confidence- access
- ungated(conversion)
- license
- Apache-2.0(recipe)+per-task
- dataset_card
- partial
Flan Collection ships 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, though, so the Apache-2.0 covers the recipe rather than the corpus itself.
- https://cdn.jsdelivr.net/gh/google-research/FLAN@main/LICENSE recorded 2026-09-29
jsDelivr copy of the LICENSE file in the google-research/FLAN GitHub repository, the full Apache License text headed "Apache License" and "Version 2.0, January 2004".
- https://cdn.jsdelivr.net/gh/google-research/FLAN@main/README.md recorded 2026-09-29
jsDelivr copy of the google-research/FLAN README, saying "The code in this repository is licensed according to the [LICENSE](LICENSE) file." and "This repository contains code to generate instruction tuning dataset collections."
- https://github.com/google-research/FLAN recorded 2026-08-13
Apache-2.0 FLAN v2 recipe
- https://huggingface.co/api/datasets/ai2-adapt-dev/flan_v2_converted recorded 2026-09-29
Hub API record for the conversion. Searched the body for cardData and license: none, and gated is false. Its description says "The original FLAN dataset needs extensive efforts to be regenerated".
- https://huggingface.co/datasets/ai2-adapt-dev/flan_v2_converted recorded 2026-09-29
HF dataset page of the conversion, showing "This is a converted version of the Flan dataset into Tulu SFT training format."
- https://huggingface.co/datasets/ai2-adapt-dev/flan_v2_converted/raw/main/README.md recorded 2026-09-29
Raw card of the Hugging Face conversion. It says "This is a converted version of the Flan dataset into Tulu SFT training format." and refers readers to the original repo "for more information about this dataset and the license".
Adoption
1 high confidenceAdoption is measured by Hugging Face downloads of the ai2-adapt-dev conversion of FLAN v2, since the canonical release lives in a GitHub repository with no download counter of its own.
- 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 confidenceStrong 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 2026-09-29
arXiv abstract of the Flan Collection paper, stating "we tease apart the effect of design decisions which enable Flan-T5 to outperform prior work by 3-17%+" across evaluation settings.
- https://arxiv.org/pdf/2301.13688 recorded 2026-09-11
Flan Collection paper full text. The per-collection comparison is in the body and in Figure 1: T5-XL-LM finetuned on Flan 2022 is set against the same model finetuned on Flan 2021, P3++ and Super-Natural Instructions, and "P3++" is defined there as all datasets in the Public Pool of Prompts. The arXiv abstract page gives the 3-17%+ headline without the per-collection rows.
- https://huggingface.co/datasets/allenai/Dolci-Instruct-SFT recorded 2026-09-29
Dolci-Instruct-SFT dataset page listing FLAN v2 with "89,981 prompts (Longpre et al., 2023)" among its sources.
Verified 2026-09-27