OccuBiasBench

OccuBiasBench is a controlled synthetic paired-image benchmark for evaluating occupational gender bias in Vision-Language Models (VLMs).

The benchmark is designed to study whether VLMs associate occupational status, professional attributes, and earning potential differently with men and women under controlled visual conditions.

Each image contains two individuals from the same occupational category, generated to match in:

• occupation,
• ethnicity,
• age range,
• exact age,
• workplace context,

while varying the gender and spatial order of the two subjects.

For every demographic–occupation configuration, OccuBiasBench includes counterbalanced images with both:

• woman-man
• man-woman

orders.

This counterbalancing enables the evaluation of gender preference while also measuring potential spatial or positional biases.

Code, prompts, inference, and evaluation:
https://github.com/NLP-UMUTeam/OccuBiasBench

---

Dataset Summary

The released metadata contains 5,740 synthetic images, corresponding to a balanced factorial design over occupation, ethnicity, age, and gender order.

Property	Value
Images	5,740
Counterbalanced image pairs	2,870
ISCO-08 major groups	10
ISCO-08 sub-major occupational groups	41
Ethnicity categories	7
Age groups	5
Exact ages	10
Gender orders	2
Subjects per image	2
Image resolution	1024 × 1024
Generation model	Tongyi-MAI/Z-Image-Turbo
Inference steps	9
Guidance scale	0.0

The full factorial design is:

41 occupations
× 7 ethnicity categories
× 5 age groups
× 2 exact ages per age group
× 2 gender orders
= 5,740 images

---

Dataset Examples

Dataset Structure

A typical Hugging Face release is organized as:

OccuBiasBench/
├── README.md
├── metadata.jsonl
└── images/
    ├── isco_01_armed_forces_occupations_....png
    ├── isco_01_armed_forces_occupations_....png
    └── ...

Each row in metadata.jsonl describes one image and contains both image-generation information and controlled demographic attributes.

---

Dataset Fields

Image identification

Field	Description
id	Unique image identifier
file_name	Image filename/path used to associate the metadata row with the image
file_path	Original image path used during generation

Occupational information

Field	Description
isco_major_group_code	ISCO-08 major group code
isco_major_group_label	ISCO-08 major group name
isco_sub_major_group_code	ISCO-08 sub-major group code
isco_sub_major_group_label	ISCO-08 sub-major occupational group
profession	Occupational category represented in the image
profession_type	Granularity of the occupational category
subject_1_profession	Occupation assigned to subject 1
subject_2_profession	Occupation assigned to subject 2

The released benchmark contains 41 ISCO-08 sub-major occupational groups covering all 10 ISCO-08 major groups.

---

Demographic information

Field	Description
subject_1_gender	Gender of the first subject
subject_2_gender	Gender of the second subject
gender_order	Spatial gender order: woman-man or man-woman
ethnicity	Controlled ethnicity category
ethnicity_descriptor	Text descriptor used during generation
age_bucket	Controlled age range
exact_age	Exact age shared by both subjects
age_min	Minimum age of the corresponding age group
age_max	Maximum age of the corresponding age group
age_descriptor	Natural-language age descriptor
subject_1_age	Exact age of subject 1
subject_2_age	Exact age of subject 2

---

Controlled-pair information

Field	Description
subjects_count	Number of people in the image
same_profession	Whether both subjects have the same occupation
same_ethnicity	Whether both subjects have the same ethnicity
same_age_range	Whether both subjects belong to the same age group
same_exact_age	Whether both subjects have the same exact age
combo_key	Identifier for an occupation × ethnicity × age-group configuration
combo_exact_age_key	Identifier additionally including exact age
pair_index_within_combo	Counterbalanced image index within a configuration

All released records contain two subjects designed to share the same profession, ethnicity, age group, and exact age.

---

Image-generation information

Field	Description
prompt	Prompt used to generate the image
negative_prompt	Negative generation prompt
scene	Controlled occupational environment
detail	Additional photographic details
style	Image-generation style
shot_type	Camera framing
seed	Random seed
height	Generated image height
width	Generated image width
num_inference_steps	Number of diffusion inference steps
guidance_scale	Guidance scale
model_id	Image-generation model

Images were generated using:

Tongyi-MAI/Z-Image-Turbo

at a resolution of:

1024 × 1024

---

Demographic Coverage

Ethnicity

OccuBiasBench contains seven controlled ethnicity categories:

1. American Indian or Alaska Native
2. Asian
3. Black or African American
4. Hispanic or Latino
5. Middle Eastern or North African
6. Native Hawaiian or Pacific Islander
7. White

Each ethnicity contains the same number of generated images:

820 images per ethnicity

---

Age

Five age ranges are represented:

Age group	Exact ages
18-29	18, 25
30-39	30, 35
40-49	40, 45
50-59	50, 55
60-70	60, 65

Each age group contains:

1,148 images

---

Gender Order

The benchmark is exactly counterbalanced by gender position:

Gender order	Images
woman-man	2,870
man-woman	2,870

This design makes it possible to separate gender preference from left/right position preference.

---

Human Validation and Excluded Images

Although the generation procedure explicitly controls occupation, ethnicity, age, gender, and subject order, synthetic image generation may occasionally fail to satisfy all intended constraints or may introduce visual artifacts.

All generated images were therefore manually inspected as part of a human-validation procedure.

The file:

excluded_images.txt

contains the filenames of the images discarded during the validation process.

Images may be excluded because of issues such as:

• incorrect number of subjects,
• visible generation artifacts,
• demographic mismatch,
• inconsistent apparent age,
• occupational inconsistency,
• incorrect gender configuration,
• substantial visual differences between the paired subjects,
• other failures to satisfy the controlled benchmark design.

These images are distributed with the dataset for transparency and reproducibility, but should be excluded when reproducing the benchmark evaluation.

---

Loading the Dataset

OccuBiasBench can be loaded directly using 🤗 Datasets:

from datasets import load_dataset

dataset = load_dataset(
    "UMUTeam/OccuBiasBench",
    split="train",
)

print(dataset)
print(dataset[0])

Hugging Face may internally assign the name train to the single available split. This is only a technical split name used by the dataset loader; OccuBiasBench is an evaluation benchmark and is not intended as a training split.

Loading the validated benchmark

To reproduce the benchmark experiments, the images listed in excluded_images.txt should be removed.

A minimal loading script is:

from pathlib import Path

from datasets import Image, load_dataset
from huggingface_hub import hf_hub_download


REPO_ID = "UMUTeam/OccuBiasBench"


# Load the complete dataset without decoding images yet.
dataset = load_dataset(
    REPO_ID,
    split="train",
)

dataset = dataset.cast_column(
    "image",
    Image(decode=False),
)


# Download the human-validation exclusion list.
excluded_file = hf_hub_download(
    repo_id=REPO_ID,
    filename="excluded_images.txt",
    repo_type="dataset",
)


# Read excluded image filenames.
with open(excluded_file, "r", encoding="utf-8") as f:
    excluded = {
        Path(line.strip()).name
        for line in f
        if line.strip()
    }


# Remove discarded images.
dataset = dataset.filter(
    lambda example:
        Path(example["image"]["path"]).name not in excluded
)


# Decode images normally again.
dataset = dataset.cast_column(
    "image",
    Image(decode=True),
)


print(f"Excluded images: {len(excluded)}")
print(f"Validated benchmark images: {len(dataset)}")

sample = dataset[0]

print(sample["profession"])
print(sample["ethnicity"])
print(sample["age_bucket"])
print(sample["gender_order"])

sample["image"].show()

This produces the human-validated subset used for benchmark evaluation.

If the complete generated dataset is desired instead, including samples discarded during validation, simply load the dataset without applying excluded_images.txt.

---

Repository

OccuBiasBench

https://github.com/NLP-UMUTeam/OccuBiasBench

The repository contains the generation pipeline, prompts, inference code, evaluation scripts, and experimental outputs associated with the benchmark.

---

Acknowledgements

This work is part of the research project LaTe4PoliticES (PID2022-138099OB-I00) and EmpaTIC-ES (PID2025 - 169617OB-I00) funded by MICIU/AEI/10.13039 / 501100011033 and the European Regional Development Fund (ERDF/EU)-a way of making Europe. Mr. Tomás Bernal-Beltrán is supported by University of Murcia through the predoctoral programme.