Tulu 3 SFT Mixture
Allen Institute for AISupervised fine-tuning dataset of roughly 939,000 samples combining 18 data sources, created by the Allen Institute for AI for its Tulu 3 post-training recipe. It covers conversation, instruction-following, math, code, safety and multilingual data across more than 70 languages, and underpins the Tulu 3 model family.
Some constituent subsets carry terms narrower than the mixture's own, which the card documents.
Openness
5 high confidence- license
- odc-by(mixed subset licenses incl CC-BY-NC)
- access
- ungated
- dataset_card
- present
ODC-BY licensed and ungated with a full dataset card, though some subsets carry non-commercial licenses.
- https://huggingface.co/api/datasets/allenai/tulu-3-sft-mixture recorded 2026-09-29
The Hub API metadata is public and ungated (a gated field set to false), with tags including "license:odc-by" and "multilinguality:multilingual".
- https://huggingface.co/datasets/allenai/tulu-3-sft-mixture recorded 2026-09-29
The dataset page is publicly served with a card that says "The Tulu 3 SFT mixture was used to train the" Tulu 3 model series.
- https://huggingface.co/datasets/allenai/tulu-3-sft-mixture/raw/main/README.md recorded 2026-09-29
The raw dataset card gives the license as "license: odc-by", and its note says "different licenses apply to subsets of the data. Some portions of the dataset are non-commercial."
Adoption
3 high confidenceAdoption is measured by Hugging Face downloads of the dataset over the trailing month, the only usage channel on record for the Tulu 3 SFT Mixture.
- https://huggingface.co/api/datasets/allenai/tulu-3-sft-mixture recorded 2026-08-13
36,267 downloads in the trailing 30 days for allenai/tulu-3-sft-mixture
Capability
4 high confidenceThe Tulu 3 paper reports that models post-trained with this mixture and the wider Tulu 3 recipe surpass the instruct versions of Llama 3.1, Qwen 2.5 and Mistral as well as GPT-4o-mini, and Marin 8B Instruct also uses it for supervised fine-tuning. Ai2's newer Dolci suite, not this mixture, post-trains Olmo 3-Think 32B, which Ai2 reports as the strongest fully open thinking model.
- https://arxiv.org/abs/2411.15124 recorded 2026-09-29
The Tulu 3 paper abstract reports models "surpassing the instruct versions of Llama 3.1, Qwen 2.5, Mistral, and even closed models such as GPT-4o-mini and Claude 3.5-Haiku."
- https://huggingface.co/marin-community/marin-8b-instruct recorded 2026-09-29
The Marin 8B Instruct card says it is an "SFT-only model. It was trained on the following datasets:" and links allenai/tulu-3-sft-mixture.
Verified 2026-09-27