You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a test set. To keep model scores meaningful, do not train on it, do not repost its items, and do not publish them in a form that could end up in training data.
Log in or Sign Up to review the conditions and access this dataset content.
RAI evaluation set
Frozen test items for measuring bias and safety in large language models, organized by the
protected classes of the New York State Human Rights Law. Run it with the open-source backend
(raieval, github.com/DATA-SCIENCE-CAPSTONE-RAI/backend): set hf_repo to this dataset and
release: v1.0 in config/dataset.yaml, then raieval fetch-release.
Contents of v1.0/
| stage | items | what |
|---|---|---|
mcq/ |
30,400 | BBQ-style multiple choice: 8 items per template and name pair |
open/ |
7,600 | BBG-style story prompts, two name orders per pair |
safety_ids.jsonl |
1,699 | WildGuardTest item ids and prompt hashes only (no prompt text) |
manifest.csv lists every item with its SHA-256 (raieval verify-release checks them);
categories.json gives display names; counts.md and sampling_report.json describe sampling.
The safety prompts come from WildGuardTest (Han et al., 2024), which may not be redistributed:
accept its terms on Hugging Face and run raieval fetch-safety.
Categories
| id | label | NYS class |
|---|---|---|
ny_sex |
Sex | sex |
ny_gender_identity |
Gender identity or expression | gender identity or expression |
ny_pregnancy |
Pregnancy-related condition | pregnancy-related condition |
ny_marital_status |
Marital status | marital status |
ny_familial_status |
Familial status | familial status |
ny_dv_victim_status |
DV victim status | domestic violence victim status |
ny_genetic |
Genetic characteristics | predisposing genetic characteristics |
ny_arrest_record |
Arrest record | arrest record |
ny_conviction_record |
Conviction record | conviction record |
ny_citizenship |
Citizenship or immigration | citizenship or immigration status |
ny_source_of_income |
Lawful source of income | lawful source of income (housing) |
ny_military_status |
Military status | military status |
ny_retaliation |
Retaliation | retaliation for opposing discrimination or requesting accommodation (§296(7)) |
ny_age |
Age | age |
ny_race_color |
Race or color | race; color |
ny_creed_religion |
Creed or religion | creed |
ny_disability |
Disability | disability |
ny_national_origin |
National origin | national origin |
ny_sexual_orientation |
Sexual orientation | sexual orientation |
Sources
Bias items are generated from templates written by the project team in the format of BBQ (Parrish et al., 2022) and BBG (Jin et al., 2025). Safety items reference WildGuardTest (allenai/wildguardmix, ODC-BY, gated).
- Downloads last month
- 37