---
license: cc-by-4.0
task_categories:
- text-generation
- feature-extraction
language:
- en
- es
- zh
- hi
- fa
- sw
- fr
- pt
- bn
- yo
- tl
tags:
- uncgpt
- personas
- synthetic-data
- multilingual
- neurips-2026
pretty_name: UncGPT — Persona Pool
size_categories:
- 1M<n<10M
configs:
- config_name: preview
  default: true
  data_files:
  - split: train
    path: data/v3/personas_preview_10k.parquet
- config_name: personas_full
  data_files:
  - split: train
    path: data/v3/personas.parquet
- config_name: names
  data_files:
  - split: train
    path: pools/names/pool.parquet
---

# UncGPT — Persona Pool

**1,000,069 synthetic personas** plus a **15,000-name registry** (zero gaps in the build report). The substrate for all UncGPT caregiving conversation generation.

Part of the [UncGPT NeurIPS 2026 Competition collection](https://huggingface.co/collections/Reza2kn/uncgpt-2026-neurips-competition-6a081d4a2a5645643dc999b4).

## Configs

| Config | Rows | What it is |
|---|---:|---|
| `preview` (default) | 10,000 | first 10k personas — for the HF dataset viewer to render quickly |
| `personas_full` | 1,000,069 | complete v3 persona pool (3.5 GB; too large for the in-browser viewer) |
| `names` | 15,000 | canonical name registry |

For programmatic use, load `personas_full` directly with `datasets` or `pyarrow`.

## Persona fields (23 columns)

- identifier + balance fields (gender, language, culture, age band)
- geographic fields (region, country)
- serialized persona JSON payload — includes Schwartz values, HEXACO traits, family / social configuration, employment, household, coping style, and stage-2 generated per-persona sub-pools (crisis profiles, symptom sets, trigger sets, value expressions, natural tendency sets)

## Other files

| Path | What it is |
|---|---|
| `data/v7/personas.parquet` | newer v7 persona snapshot |
| `pools/static/*.yaml` | reusable seed pools (symptoms, triggers, tendencies, employments, households, social networks, summary variants) |
| `pools/generated/gold/*.json` | curated gold pools used during generation |
| `pools/generated/v4_anticollapse/personas.parquet` | anti-collapse persona experiment |
| `metadata/v3/batch.json` | generation batch metadata + audits |
| `metadata/v3/pools_manifest.json` | index of uploaded diversity pools |
| `pools/names/build_report.json` | names pool build report (gaps: `[]`) |
| `figures/v3/semantic/similarity_noncollapse_summary.json` | numerical non-collapse summary |

## Provenance

All personas are **synthetically generated**. No human data was scraped. Generation pipeline: [`Reza2kn/KakoVerse`](https://github.com/Reza2kn/KakoVerse). Seed lexicons are publicly licensed (OLDI Wikipedia MT seeds, Common Voice 24 fa for orthographic priors).

## Semantic diversity

See semantic atlas + density / centroid charts in the [companion GitHub repo](https://github.com/Reza2kn/uncgpt-2026-neurips/tree/main/data/personas).

## License

CC-BY-4.0.