PersonaHub
Tencent AI LabPersonaHub is a persona-driven synthetic data resource from Tencent AI Lab, backed by the paper 'Scaling Synthetic Data Creation with 1,000,000,000 Personas' (arXiv 2406.20094). The repo releases a large set of distilled personas (200K preview plus ~370M elite personas) plus seed synthetic samples such as 50K math problems, 50K logical-reasoning problems, and 50K instructions. It is a method and dataset for generating diverse, scalable synthetic instruction and reasoning data.
Verified live 2026-06-22 via primary sources. Publicly downloadable and ungated under CC-BY-NC-SA-4.0, which permits redistribution for non-commercial research.
Openness
2 high confidence- license
- cc-by-nc-sa-4.0(non-commercial)
- card
- present
- ungated
- yes
- scope
- ~370M personas+seed samples
Publicly downloadable and ungated, and licensed CC-BY-NC-SA-4.0, which permits redistribution for non-commercial research and forbids commercial use. Moved from 5/open on 2026-08-01 by the universal license scale. The map had answered the non-commercial question two ways: `command-r` is 2/restricted under CC-BY-NC while this was 5/open under CC-BY-NC-SA, on the reasoning that a corpus's openness question is redistributability rather than commercial reuse. The scale settles it on the commercial-use test, and the Open Definition agrees - data is not open unless it may be reused commercially. `restricted` is new to the dataset class vocabulary for this: it was the only product type with no word between `open` and `gated`.
- https://huggingface.co/datasets/proj-persona/PersonaHub recorded 2026-06-22
CC-BY-NC-SA-4.0 license, ungated card, persona counts and paper reference
Adoption
2 high confidencePublicly downloadable and ungated, and licensed CC-BY-NC-SA-4.0, which permits redistribution for non-commercial research and forbids commercial use. Moved from 5/open on 2026-08-01 by the universal license scale. The map had answered the non-commercial question two ways: `command-r` is 2/restricted under CC-BY-NC while this was 5/open under CC-BY-NC-SA, on the reasoning that a corpus's openness question is redistributability rather than commercial reuse. The scale settles it on the commercial-use test, and the Open Definition agrees - data is not open unless it may be reused commercially. `restricted` is new to the dataset class vocabulary for this: it was the only product type with no word between `open` and `gated`.
- https://huggingface.co/api/datasets/proj-persona/PersonaHub recorded 2026-07-04
downloads field = 7,395 (30-day)
Capability
2 high confidencePublicly downloadable and ungated, and licensed CC-BY-NC-SA-4.0, which permits redistribution for non-commercial research and forbids commercial use. Moved from 5/open on 2026-08-01 by the universal license scale. The map had answered the non-commercial question two ways: `command-r` is 2/restricted under CC-BY-NC while this was 5/open under CC-BY-NC-SA, on the reasoning that a corpus's openness question is redistributability rather than commercial reuse. The scale settles it on the commercial-use test, and the Open Definition agrees - data is not open unless it may be reused commercially. `restricted` is new to the dataset class vocabulary for this: it was the only product type with no word between `open` and `gated`.
- https://arxiv.org/abs/2406.20094 recorded 2026-07-04
Qwen2-7B on ~1.07M persona math problems hits 64.9% MATH, rivaling GPT-4-turbo-preview
- https://huggingface.co/datasets/allenai/tulu-3-sft-personas-math recorded 2026-07-04
Tulu 3 conditions on ~250K PersonaHub personas
Unchanged since 2026-08-01 (last edited, not re-checked)