datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
llm-bias-detection
LLM Bias Detection Evaluation Traces
Evaluation data accompanying Navigating the digital spectrum: Assessing
political bias, stability, and downstream fairness in Large Language Models
(arXiv:2609.08637).
Licence scope: CC BY 4.0 covers the authors' original documentation,
templates, selection/arrangement and author-generated tables. It does not
relicense source text or annotations. IBM retains CC BY-SA 3.0; hate-corpus
components retain CC BY 4.0, CC0 or MIT as documented in… See the full description on the dataset page: https://huggingface.co/datasets/nishan-chatterjee/llm-bias-detection.LLM_Bias_EECS182_Project_New_DatasetLLM_Bias_EECS182_Projectllm-bias-mt-benchThere are some dubious responses in answers.json that need inspection.
To find them out, run:
import json
DUBIOUS_MARKER = "[ERROR: Failed to generate answer"
def is_dubious(answer_text: str) -> bool:
return not answer_text or DUBIOUS_MARKER in answer_text
with open("answers.json") as f
answers = json.load(f)
dubious_indices = [
idx for idx, answer in enumerate(answers) if is_dubious(answer.get("answer_text", ""))
]
These answers are mostly either null text or has this:… See the full description on the dataset page: https://huggingface.co/datasets/sayakpaul/llm-bias-mt-bench.geo-bias-llm
Geo Bias LLM — Institutional Prestige as Geographic Bias in Large Language Models
Experimental data for the paper "Institutional Prestige as Geographic Bias in
Large Language Models: Evidence from Three Factorial Experiments with Bootstrap
Confidence Intervals" (Leyva-Vázquez, 2026).
Paper page: https://huggingface.co/papers/2608.18107
Code: https://github.com/mleyvaz/geo-bias-llm
Author: Maikel Leyva-Vázquez — Universidad Bolivariana del Ecuador /
Universidad de Guayaquil… See the full description on the dataset page: https://huggingface.co/datasets/mleyvaz/geo-bias-llm.LLM_Bias_Detection_DatasetLLM Bias Detection Dataset README
Overview
This dataset (entries B1–B1000) is designed to train and evaluate AI models for detecting biases in large language model (LLM) outputs, focusing on cybersecurity, tech professions, and social contexts. It aligns with NIST AI Risk Management Framework and OWASP Top 10 for LLMs to identify gender, racial, geographic, and socioeconomic biases.
Dataset Purpose
Objective: Enable ML engineers and cybersecurity professionals to train models to detect and… See the full description on the dataset page: https://huggingface.co/datasets/darkknight25/LLM_Bias_Detection_Dataset.LLM-Bias4-Eval
📊 LLM-Bias4-Eval
LLM-Bias4-Eval is a biased dataset for analyzing structural biases in LLM evaluations. It aims to expose how language models can exhibit unintended preferences in structured decision-making tasks. Specifically, it targets four types of biases: position, bandwagon, verbosity, and sentiment.
📂 Dataset Files
The dataset consists of three files, each focusing on specific bias types and domains:
File Name
Bias Type(s)
Description… See the full description on the dataset page: https://huggingface.co/datasets/joyfine/LLM-Bias4-Eval.phishing-llm-bias-audit
LLM Phishing-Vulnerability Bias Audit Dataset
A multi-provider empirical dataset capturing how 14 open-source LLM configurations (across 5 inference providers) select which of three generated personas is "most vulnerable to phishing." 855 persona records / 285 forced-choice workflows.
Important. This dataset is about LLM behaviour under controlled prompts, not about real-world phishing susceptibility of any demographic group. Selecting a persona as "vulnerable" is the LLM's choice;… See the full description on the dataset page: https://huggingface.co/datasets/Julia569922/phishing-llm-bias-audit.llm-nationality-bias-global-narratives
Representational Harms in Global LLM Narratives: Nationality Bias Dataset
Dataset Summary
This dataset contains 292,500 LLM-generated narratives across 195 globally-recognized nations, created to examine how national identity cues in prompts shape narrative content and representation. Generated using GPT-4.1 Nano, the dataset systematically varies the nationality of dominant characters across power-laden scenarios in Learning, Labor, and Love domains. This enables… See the full description on the dataset page: https://huggingface.co/datasets/ilana27/llm-nationality-bias-global-narratives.llm-nationality-bias-us-narratives
Representational Harms in US-Based LLM Narratives: Nationality Bias Dataset
Dataset Summary
This dataset contains 9,710 LLM-generated narratives that reference non-US national identities, extracted from a larger corpus of 500,000 stories generated by GPT-3.5, GPT-4, Claude 2.0, Llama 2, and PaLM 2. The stories were generated in response to open-ended prompts set in US contexts across three domains: Learning, Labor, and Love. This dataset was created to study… See the full description on the dataset page: https://huggingface.co/datasets/ilana27/llm-nationality-bias-us-narratives.Prompt-Framing-Bias-LLM
🧠 Prompt Framing and Socio-Demographic Bias in LLMs
Dataset by: Lipi ChandrakarMSc Data Science and Analytics, University of Hertfordshire (2025)
📄 Associated Paper
More Responses, More Harm: How Prompt Framing and Decoding
Choices Shape Socio-Demographic Bias in Open-Source LLMs📎 https://doi.org/10.5281/zenodo.20169051
If you use this dataset please cite:Chandrakar, L. (2025). More Responses, More Harm.
Zenodo. https://doi.org/10.5281/zenodo.20169051
📘… See the full description on the dataset page: https://huggingface.co/datasets/lipichandrakar/Prompt-Framing-Bias-LLM.llm-drug-bias
LLM Medication Selection Bias Dataset - Input Prompts
Overview
This dataset contains the input prompts used in a study analyzing biases in drug recommendations made by large language models (LLMs). The dataset includes clinical vignettes with various patient demographics to examine how LLMs respond to different patient characteristics when recommending medications.
Data File
drug_bias_prompts.xlsx
This file contains the standardized clinical… See the full description on the dataset page: https://huggingface.co/datasets/Medical-AI-Lab/llm-drug-bias.racial-bias-llm
Color-Blind Bias in Frontier Language Models
This repository contains the data, code, and metadata accompanying the paper
"Color-Blind Bias in Frontier Language Models: An Evaluation Framework with
Validated Racial Attitude Scales", under double-blind review at NeurIPS 2026
(Evaluations & Datasets Track).
Submission status: anonymous for review.
What this evaluation measures
We applied six validated human racial-attitude scales to 32 frontier language
models from 8… See the full description on the dataset page: https://huggingface.co/datasets/latenttensor26/racial-bias-llm.LLM_Select_Bias