FairStress

Dataset for the paper "Asymmetrically Unfair Decision Making and Reasoning in LLMs".

FairStress tests LLM fairness as a decision process rather than a single answer. It includes 300 decision-making scenarios across 29 high-stakes domains (hiring, healthcare, law, housing, etc.), each posed with two candidates who differ by a protected-attribute contrast (race, gender, religion, disability, age, ...). Every scenario is instantiated at 4 degrees of demographic-identity signaling (implicit to explicit) and, at the robustness level, paired with one of 78 adversarial prefixes that argue for one candidate or the other.

FairStress gives the complete, full analysis of procedural bias — not just whether a model's final answer is correct, but whether it holds a correct decision, and a tied decision, equally robustly regardless of which group it favors.

About the work

Standard bias benchmarks (BBQ, StereoSet, CrowS-Pairs) score models on the answer alone. FairStress instead measures procedural fairness — whether the process that produced a decision is fair — at four levels: (1) the decision itself, (2) how firmly the model holds that decision under an adversarial prefix, (3) whether its stated reasoning is biased, and (4) whether its internal activations are. Each of the 300 base scenarios (two candidates, one domain, one decision) is rendered as a ground-truth version (one candidate is objectively stronger) and an ambiguous version (a true tie), across 33 minority/majority identity contrasts, 4 degrees of how explicitly that identity is signaled, and up to 78 adversarial prefixes arguing for one candidate or the other.

Headline statistics

Quantity	Value
Base scenarios	300
Domains	29
Demographic groups	31
Minority–majority contrasts	33, in 8 categories
Intervention degrees	4 (D0 no identity → D3 explicit label)
Adversarial prefixes	78, in 15 families, 3 classes

Domains (29 domains with base-scenario counts — every domain is instantiated across every applicable contrast/degree/prefix)

Domain	n	Domain	n
Hiring & Employment	20	Insurance Claims	10
Medical & Healthcare	18	Moral Tradeoffs	8
Legal / Criminal Justice	18	Environmental Justice	8
Social Services	14	Content Moderation	8
Financial Decisions	12	Education (Student-facing)	8
Disability Accommodation	10	Data Privacy	8
Mental Health	10	Consumer Protection	8
Immigration & Asylum	10	Legal Records	8
Education	10	Housing (Tenant-facing)	8
Public Safety	10	Epistemic Credibility	8
Veterans Benefits	10	Narrative / Personal Stmt.	6
Criminal Reentry	10
Housing & Tenancy	10
Academic Integrity	10
Child & Family	10
Child & Family (Application)	10
Workplace Discipline	10
Elder Care	10
Total			300

Contrast categories (33 contrasts, 31 demographic groups)

Category	n	Example contrasts
Race	4	Black/white, Hispanic/white, Arab/Hispanic, Arab/Black
Sex / gender	7	Black female/male, white trans woman/man, white trans man/man
Sexuality	2	gay/heterosexual, lesbian/heterosexual
Race × sex	7	Black female/white male, Hispanic trans woman/white male
Religion	8	Jewish/Muslim, Muslim/Christian, Sikh/Christian, Hindu/Muslim
Religion × gender	3	atheist male/Muslim female, Jewish male/Christian female
Age	1	older/younger worker
Disability	1	disabled/non-disabled
Total	33

Adversarial prefixes (78, 15 families, 3 classes)

Class	Families	n
Directional (matched pro-A/pro-B pairs)	authority, social consensus, representation, historical guilt, synthetic evidence, emotional appeal, legal liability, synthetic statistics, disadvantage claim	36
Neutral	evaluation awareness, stakes framing, reactance	18
Control	neutral-relevant, neutral-irrelevant, placebo	24
Total		78

Splits

Split	What it is
full	Recommended split to run. The complete generated corpus — every scenario × domain × signaling-degree × adversarial-prefix combination. This is the exact data behind the paper's results; running a model against it gives the complete, exact analysis of procedural bias.
core	FairStress-Core — a subset of full, selected via 2PL IRT (Fluid Benchmarking) rather than randomly. A lighter, cheaper option when running the full 13M-row corpus isn't feasible.

Usage

from datasets import load_dataset

# full corpus — recommended: the complete, exact analysis of procedural bias
full = load_dataset("PardisSzah/fairstress", split="full")

# FairStress-Core — lighter/cheaper subset
core = load_dataset("PardisSzah/fairstress", split="core")