Tencent AI Lab

license

cc-by-nc-sa-4.0(non-commercial)

card

present

ungated

yes

scope

~370M personas+seed samples

PersonaHub is publicly downloadable and ungated, licensed CC-BY-NC-SA-4.0, which permits non-commercial reuse but forbids commercial use. Because the Open Definition treats non-commercial-only data as not fully open, that leaves PersonaHub between open and gated.

• https://huggingface.co/api/datasets/proj-persona/PersonaHub recorded 2026-09-29

  Hub API record for proj-persona/PersonaHub, listing the tag "license:cc-by-nc-sa-4.0" and a description of "a collection of 1 billion diverse personas".

• https://huggingface.co/datasets/proj-persona/PersonaHub recorded 2026-09-29

  PersonaHub dataset page, public with no access-request gate (its embedded page data reports gated false), showing the license "cc-by-nc-sa-4.0" and the paper "Scaling Synthetic Data Creation with 1,000,000,000 Personas".

• https://huggingface.co/datasets/proj-persona/PersonaHub/raw/main/README.md recorded 2026-09-29

  Raw card of PersonaHub with a YAML header stating "license: cc-by-nc-sa-4.0" and a data release section listing "370,000,000 elite personas (added in Feb 2025)" and "200,000 personas (early preview)".

Adoption is measured by Hugging Face downloads of the dataset over the trailing month, the only usage channel on record for PersonaHub.

• https://huggingface.co/api/datasets/proj-persona/PersonaHub recorded 2026-08-13

  9,346 downloads in the trailing 30 days for proj-persona/PersonaHub

PersonaHub's headline result is a Qwen2-7B model fine-tuned on roughly 1.07 million persona-generated math problems, reaching about 65% on MATH and matching GPT-4-turbo-preview. Its impact runs mostly through datasets built on its method, such as Tulu 3's persona-math mixture, rather than through direct large-scale use of the personas themselves, unlike SmolTalk, FineWeb-Edu and DCLM-Baseline.

• https://arxiv.org/html/2406.20094 recorded 2026-09-29

  Paper HTML says "our approach allows a 7B LLM to achieve 65% on MATH" and describes fine-tuning Qwen2-7B on "1.07 million math problems and evaluate its greedy decoding outputs on the above two test sets".

• https://huggingface.co/datasets/allenai/tulu-3-sft-personas-math recorded 2026-09-29

  Tulu 3 persona-math dataset page whose card says "we expand the methodology in" the PersonaHub paper "by using personas" and lists "149960" examples.