Dolci
Allen Institute for AIAi2's post-training data suite for Olmo 3, with separate mixes for supervised fine-tuning, DPO preference tuning and reinforcement learning with verifiable rewards (RLVR) across the Instruct and Think variants. The mixes combine existing sets such as OpenThoughts 3, FLAN v2, OpenAssistant, Aya, Tulu 3 persona data and WildChat with new Ai2 prompts for precise instruction following, Python algorithms, logic puzzles and tool use.
A family record covering the Dolci-Instruct and Dolci-Think SFT, DPO and RL sets.
Openness
5 high confidence- license
- ODC-BY
- gated
- false
- dataset_card
- present
ODC-BY, ungated, per-stage dataset cards.
- https://huggingface.co/api/datasets/allenai/Dolci-Instruct-SFT recorded 2026-09-29
The Hub API record for allenai/Dolci-Instruct-SFT carries the card text "licensed under ODC-BY. It is intended for research and educational use". Searched the body for gated auto and gated manual values: none, the gated field is false.
- https://huggingface.co/datasets/allenai/Dolci-Instruct-SFT recorded 2026-09-29
The dataset page renders the card note "licensed under ODC-BY. It is intended for research and educational use" and shows the file viewer with no gate. Searched the body for Access gated dataset, Request access and You need to agree: none.
- https://huggingface.co/datasets/allenai/Dolci-Instruct-SFT/raw/main/README.md recorded 2026-09-29
The raw dataset card front matter declares "license: odc-by" and its body states "It contains 2,152,112 samples from the following sets".
Adoption
2 high confidenceAdoption is measured by Hugging Face downloads of allenai/Dolci-Instruct-SFT, Dolci's only tracked usage channel.
- https://huggingface.co/api/datasets/allenai/Dolci-Instruct-SFT recorded 2026-08-13
3,807 downloads in the trailing 30 days for allenai/Dolci-Instruct-SFT
Capability
5 high confidenceAi2's launch post reports Olmo 3-Think 32B, post-trained on Dolci, as the strongest fully open thinking model it is aware of, while noting it still trails some open-weight models such as Qwen 3 32B. The comparison is between finished models rather than an ablation of Dolci against other mixes.
- https://allenai.org/blog/olmo3 recorded 2026-09-29
The Ai2 Olmo 3 launch post says "Dolci provides separate mixes for each stage of post-training: SFT, DPO, and RLVR" and its benchmark table lists Olmo 3-Think (32B) at MATH "96.1", HumanEvalPlus "91.4" and IFEval "89.0", concluding "Think 32B emerges as the strongest fully open thinking model."
Verified 2026-09-27