README.md

---
license: cc-by-4.0
pretty_name: ProfileFoundry-Core-100K
configs:
- config_name: person_objects
  default: true
  data_files:
  - split: train
    path: person_objects.parquet
- config_name: profiles_jsonl
  data_files:
  - split: train
    path: profiles.jsonl
- config_name: profiles
  data_files:
  - split: train
    path: profiles.parquet
- config_name: addresses
  data_files:
  - split: train
    path: addresses.parquet
- config_name: employment
  data_files:
  - split: train
    path: employment.parquet
- config_name: education
  data_files:
  - split: train
    path: education.parquet
- config_name: social_handles
  data_files:
  - split: train
    path: social_handles.parquet
- config_name: health_allergies
  data_files:
  - split: train
    path: health_allergies.parquet
- config_name: events
  data_files:
  - split: train
    path: events.parquet
- config_name: households
  data_files:
  - split: train
    path: households.parquet
- config_name: employers
  data_files:
  - split: train
    path: employers.parquet
- config_name: relationships
  data_files:
  - split: train
    path: relationships.parquet
---

# ProfileFoundry-Core-100K

Synthetic Person Object dataset produced by ProfileFoundry. This is the
vetted `profilefoundry-core` v1.0 release package; when uploaded to Hugging
Face, verify the remote `MANIFEST.json` against this package before citing the
hosted artifact as current.

* SDK version: **0.1.0.dev0**
* Generation date: **2026-05-24**
* Exported at: **2026-05-24T00:00:00Z**
* Total profiles: **100,000**
* Total events: **709,228**
* Total social handles: **135,103**
* Total health allergies: **36,792**
* Total households: **40,338**
* Total employers: **52,491**
* Total relationships (edges): **518,564**

## Locale mix

- `AU`: 7,000 profiles
- `CA`: 8,000 profiles
- `IE`: 4,000 profiles
- `IN`: 20,000 profiles
- `NZ`: 3,000 profiles
- `PH`: 3,000 profiles
- `UK`: 20,000 profiles
- `US`: 35,000 profiles

## Files

| File | Contents |
|---|---|
| `profiles.jsonl` | Canonical Person Object, one JSON per line |
| `person_objects.parquet` | Viewer-friendly complete Person Object rows with nested sections encoded as JSON strings |
| `profiles.parquet` | Flat snapshot (one row per profile) |
| `addresses.parquet` | Current and historical move addresses with valid_from/to and source_event_ids |
| `employment.parquet` | One row per (profile, employment) with FK employer_id |
| `education.parquet` | One row per (profile, education record) |
| `social_handles.parquet` | One row per emitted social handle with platform and rendered handle |
| `health_allergies.parquet` | One row per emitted health allergy |
| `events.parquet` | Long-format event timeline with typed payload columns and payload_json |
| `households.parquet` | One row per household with composition summary, address fields, and member_profile_ids_json |
| `employers.parquet` | One row per unique employer_id with pool or inferred industry/size metadata, distinct/current employee counts, and employee profile ID lists |
| `relationships.parquet` | Bidirectional graph edges across families, cohabiting partners, non-family household co-members, and a capped employer-colleague sample; rows carry household context and colleague rows carry employer_id |
| `MANIFEST.json` | File inventory with SHA-256 hashes and row counts |
| `dataset_card.md` | This file |

## License

* Dataset: CC-BY-4.0
* SDK: MIT
* Embedded reference data retains its upstream license.

## Provenance

Every profile carries `generation.global_seed`, `generation.profile_seed`,
`generation.sdk_version`, and `generation.reference_manifest_hash`, which
together identify the deterministic generation context. Re-running with the
same `--seed` and pinned generation date reproduces the generated content
including `generation.generated_at`.

## Validation

Per-locale KS gaps (smaller = closer to reference data) and consistency-rate.

| Locale | age_ks_M | age_ks_F | edu_ks | marital_ks | consistency |
|---|---|---|---|---|---|
| US | 0.104 | 0.100 | 0.004 | 0.091 | 100.00% |
| UK | 0.101 | 0.093 | 0.011 | 0.095 | 100.00% |
| IN | 0.124 | 0.124 | 0.002 | 0.104 | 100.00% |
| CA | 0.109 | 0.102 | 0.007 | 0.080 | 100.00% |
| AU | 0.116 | 0.102 | 0.010 | 0.083 | 100.00% |

Release target: KS <= 0.10 per attribute and mean <= 0.07. This v1.0 release discloses the observed gaps rather than tuning them to force the target.

## Leakage audit

| Audit | Profiles checked | Findings | Mode |
|---|---|---|---|
| self_collision | 100000 | 1045 | exact |
| wikidata | 100000 | 342 | bloom |
| email | 100000 | 0 | syntax_only |

Wikidata audits use a Bloom filter (FP rate ~1e-4); reported leak counts are upper bounds.

## Bias and limitations

* Reference distributions are 2021–2023 census/ACS bootstraps; living
  populations drift, and the dataset does not reflect post-2023 changes.
* Cross-household family edges (extended family) are deferred to v1.1.
* `households.composition_type_inferred` is derived from represented family
  edges and member ages, not from the original generator spec. Because v1 is
  adults-only, `*_with_children` means represented adult children.
* `employment.parquet` keeps the current employer first and historical jobs
  after it with `valid_to` populated.
* Phone and address fields are generated synthetically and should be treated
  as test data, but plausible strings can still coincide with real-world
  records by chance. Release leakage evidence is published in the leakage
  section.
* Income, net worth, and credit scores are statistical realisations, not
  individual-truth values. Don't use them for downstream decisions about
  real people.
* This release contains structured Person Objects only. It does not include
  auxiliary generated-content datasets.