| --- |
| license: cc-by-4.0 |
| language: |
| - vi |
| pretty_name: GreenPersona Dataset |
| size_categories: |
| - 10K<n<100K |
| task_categories: |
| - token-classification |
| - text-generation |
| tags: |
| - synthetic |
| - pii |
| - persona |
| - vietnamese |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-*.parquet |
| - split: validation |
| path: data/validation-*.parquet |
| - split: test |
| path: data/test-*.parquet |
| --- |
| |
| # GreenPersona Dataset |
|
|
| GreenPersona is a **fully synthetic** set of 97,879 Vietnamese personal profiles ("person vault", version 0.3.0). Each record is one fictional person with a consistent set of personal data: demographics, name, date of birth, residence and birthplace, national ID (CCCD), passport, driver licence, vehicles, phone numbers, emails, online accounts, bank accounts, payment cards, and sensitive attributes (ethnicity, religion, political affiliation, health, legal record, financial profile). |
|
|
| The field taxonomy follows Vietnam's Law on Personal Data Protection (91/2025/QH15) and Decree 356/2025/NĐ-CP: basic personal data (Article 3) and sensitive personal data (Article 4). The `Điều khoản` column in [DATASHEET.md](DATASHEET.md) maps each field to its clause. |
|
|
| The dataset is intended as a source of identities for generating and evaluating PII detection, redaction and guardrail systems. |
|
|
| ## Splits |
|
|
| | Split | Persons | |
| |---|---| |
| | train | 78,221 | |
| | validation | 9,798 | |
| | test | 9,860 | |
|
|
| The split is assigned by hashing `person_id`, so a given person appears in only one split. |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("GreenNode/GreenPersona-Dataset") |
| p = ds["train"][0] |
| print(p["full_name"], p["id"], p["phone_numbers"]) |
| ``` |
|
|
| ## Schema |
|
|
| One row per person, 91 flat columns. Single-valued fields are scalar columns; a missing passport, driver licence, etc. is `null`. |
|
|
| A person can have several phones, emails, online accounts, bank accounts, cards or vehicles. Each of those groups is stored as **parallel list columns**: element `i` of every column in a group describes the same item. For example, `card_numbers[1]`, `card_expiry[1]` and `card_cvv[1]` all belong to the person's second card. A person with none has empty lists. |
|
|
| | Group | Columns | |
| |---|---| |
| | Demographics (seed) | `sex`, `age`, `age_band`, `birth_cohort`, `marital_status`, `education`, `occupation`, `skills_and_expertise` (list), `hobbies` (list) | |
| | Residence | `province_code`, `province`, `zone` (34-province administrative map from 2025) | |
| | Birth registration | `birth_province_legacy`, `birth_province_legacy_cccd_code`, `birth_province_code`, `birth_province` | |
| | Name | `full_name`, `surname`, `middle_name`, `given_name`, `name_prefix` (e.g. `Y`, `H'`), `name_style` | |
| | Birth, nationality | `dob`, `nationalities`, `nationality_codes` | |
| | National ID (CCCD) | `id`, `id_card_type`, `id_issued_on`, `id_expires_on`, `id_issued_by` | |
| | Passport | `passport`, `passport_issued_on`, `passport_expires_on` | |
| | Driver licence | `driver_license`, `driver_license_classes`, `driver_license_issued_on` | |
| | Vehicles (parallel lists) | `license_plates`, `plate_codes`, `vehicle_types`, `vehicle_registered_on`, `vehicle_registration_province_codes`, `vehicle_registration_provinces` | |
| | Phones (parallel lists) | `phone_numbers`, `phone_carriers`, `phone_is_primary` | |
| | Emails (parallel lists) | `emails`, `email_domains`, `email_is_primary` | |
| | Online accounts (parallel lists) | `social_platforms`, `social_usernames`, `social_linked_phones` | |
| | Bank accounts (parallel lists) | `bank_codes`, `bank_names`, `bank_numbers`, `bank_account_names`, `bank_alias_of`, `bank_opened_on` | |
| | Payment cards (parallel lists) | `card_numbers`, `card_bins`, `card_bank_codes`, `card_schemes`, `card_network_types`, `card_types`, `card_linked_accounts`, `card_holder_names`, `card_issued_on`, `card_expiry`, `card_cvv`, `card_is_test_bin` | |
| | Ethnicity, religion, politics | `ethnic`, `religion`, `party_member`, `party_joined_year`, `civic_engagement` | |
| | Health | `icd10` (chronic conditions, ICD-10 codes), `blood_type` | |
| | Legal record | `legal_record`, `offense_type`, `offense_year` | |
| | Financial profile | `income_band`, `monthly_income_vnd`, `loans`, `debt`, `debt_group`, `bad_debt`, `credit_grade`, `insurance_products`, `has_securities_account` | |
| | Meta | `person_id`, `split`, `reserved_fields` (fields generated in reserved ranges) | |
|
|
| The columns are a flattened view of the nested vault schema. [DATASHEET.md](DATASHEET.md) documents each original field (nested path such as `documents.national_id.number` ↔ column `id`): type, nullability, dependencies, legal clause, and the status of each reference table. [vault_fields.yaml](vault_fields.yaml) is the machine-readable nested schema. [manifest.json](manifest.json) records the build hashes and generator versions. |
|
|
| ## How it was generated |
|
|
| 1. **Seed:** 7 structured columns (sex, age, marital status, education, occupation, skills, hobbies) from [nvidia/Nemotron-Personas-Vietnam](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Vietnam). Implausible combinations are filtered out. Free-text persona descriptions and the original location columns are dropped. |
| 2. **Sampling:** the remaining fields are sampled from reference tables and probabilistic graphical models conditioned on the seed attributes. Examples: ethnicity and religion by province; name style by ethnicity; national ID digits consistent with sex, birth year and birth province; income and debt by occupation, education and zone. |
| 3. **Validation:** cross-field consistency rules are checked for every record. This build passed with zero violations. |
|
|
| ## Limitations and responsible use |
|
|
| - **All persons are fictional.** No record describes a real individual. Several reference tables are marked `provisional` (see DATASHEET.md), so distributions are approximations, not official statistics. |
| - This build uses `id_realism = realistic`: national IDs, bank account numbers and card numbers follow real formats, including real NAPAS BINs and valid check digits. A generated value may therefore coincide with a real identifier by chance. Do not use these values to look up, contact or transact with anyone, and do not treat a match as real personal data. |
| - Use the dataset only for research and for building or evaluating privacy-protecting systems. |
|
|
| ## License and attribution |
|
|
| Released under CC BY 4.0. The seed attributes come from [nvidia/Nemotron-Personas-Vietnam](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Vietnam) (CC BY 4.0, NVIDIA). Third-party reference tables keep the licences listed in DATASHEET.md. |
|
|
