AI Potluck
Model components / Training & synthetic datasets

PersonaHub

Tencent AI Lab

Persona-driven synthetic data resource from Tencent AI Lab, backed by the paper "Scaling Synthetic Data Creation with 1,000,000,000 Personas". It releases a large set of distilled personas - a 200,000-persona preview alongside roughly 370 million elite personas - plus seed synthetic samples including 50,000 math problems, 50,000 logical-reasoning problems and 50,000 instructions.

Tulu 3's persona-math set expands the same methodology. Verified 2026-08-13 via the proj-persona/PersonaHub dataset card on Hugging Face.

Openness

2 high confidence
2.0
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. Openness turns on the commercial-use test rather than on redistributability alone, and the Open Definition agrees: data is not open unless it may be reused commercially. That leaves this corpus between open and gated, which is what restricted records.

Adoption

2 high confidence
2.0

9,346 downloads in the trailing 30 days for proj-persona/PersonaHub, which bands at 1K-10K, level 2 on the dataset adoption scale.

Capability

4 medium confidence
4.0

The paper's headline training result is the basis: Qwen2-7B fine-tuned on ~1.07M persona-generated math problems reaches 64.9% on MATH, rivaling GPT-4-turbo-preview at the time. Persona Hub is 1 billion personas curated from web data and used to synthesize mathematical and logical reasoning problems at scale, and the lineage carries downstream: Tulu 3's SFT persona-math set conditions on ~250K of these personas and credits the persona methodology for its 149,960 examples, so the corpus is an ingredient in a mixture already scored 4 here. It sits at 4 rather than 5 because the 5s in this category - SmolTalk, FineWeb-Edu, DCLM-Baseline - both lead head-to-head ablations and are the default others adopt wholesale, whereas PersonaHub is a generation method plus seed samples whose adoption runs through derivative sets. That is the same shape as tulu-3-sft-mixture and fineweb, both also 4.

Verified 2026-08-13