metadata

license: cc-by-nc-sa-4.0
task_categories:
  - text-classification
language:
  - en
tags:
  - legal
  - bias-detection
  - robustness
size_categories:
  - 10K<n<100K
pretty_name: RobustBiasBench
dataset_preview: data/robustbiasbench_dataset.csv
dataset_info:
  features:
    - name: id
      dtype: int64
    - name: date
      dtype: string
    - name: bias_type
      dtype: string
    - name: normative_framing
      dtype: string
    - name: source
      dtype: string
    - name: policy
      dtype: string
    - name: bias_type_merged
      dtype: string
  download_size: 5300000
  dataset_size: 5300000
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/robustbiasbench_dataset.csv

RobustBiasBench Dataset Description

This document provides an overview of the features and labels included in the RobustBiasBench dataset, which consists of over 18,000 policy excerpts annotated for bias type and normative framing.

📄 Dataset Format

The dataset is stored in CSV format and contains the following columns:

Column Name	Description
`id`	Unique numeric ID for each policy excerpt
`date`	Year of the policy (extracted from the source document)
`bias_type`	Fine-grained original bias label assigned by annotators (e.g., `age`, `gender`, `citizenship`)
`normative_framing`	Whether the bias is presented implicitly or explicitly (`implicit`, `explicit`, or `no_bias`)
`source`	URL or citation source for the original policy document
`policy`	Text excerpt of the policy (typically 1–3 sentences)
`bias_type_merged`	Mapped bias class used for evaluation: one of `group_1`, `group_2`, or `no_bias`

🏷️ Label Description

`bias_type`

Original annotation capturing specific bias domains:

age, gender, race/culture, religion, disability → Group 2 (Demographic Bias)
economic, education, political, citizenship, criminal_justice → Group 1 (Systemic Bias)
no_bias → Procedural or neutral statements

`bias_type_merged`

Mapped classes used for modeling:

group_1: Systemic/institutional bias
group_2: Demographic/identity-based bias
no_bias: Factual or operational content

`normative_framing`

Captures how bias is framed:

explicit: Bias is directly stated (e.g., "women are not eligible")
implicit: Bias is implied through structure or condition (e.g., "must be a citizen to apply")
no_bias: Used only when bias_type = no_bias

📊 Class Distribution

The dataset includes the following number of annotated examples:

No Bias: 6,017 examples
Systemic Bias (group_1): 6,246 examples
Demographic Bias (group_2): 6,141 examples

This balance ensures that the model does not overfit to any single category and can learn to differentiate across nuanced cases of policy bias.

🔒 License

This dataset is released under the CC BY-NC-SA 4.0 License and is intended for academic use in fairness and robustness research.