license: other
license_name: "original-license-preserving-aggregation-with-per-example-source-licenses"
language:

• en
  task_categories:
• text-classification
  size_categories:
• 10K<n<100K
  tags:
• benchmark
• evaluation
• bias-detection
• policy-analysis
  pretty_name: "BRIDGE: Bias Register-Integrated Dataset for Generalized Evaluation (Anonymous Submission)"

BRIDGE: Bias Register-Integrated Dataset for Generalized Evaluation (Anonymous Submission)

Dataset Description

BRIDGE is a 3-way classification benchmark for policy-analysis style text, where each example is labeled as harmful, harmless, or antibias. The dataset is designed to test whether models can distinguish harmful framing, explicit anti-bias language, and neutral procedural text across different source types, eras, and registers.

This card is written for an anonymous NeurIPS 2026 Evaluations & Datasets submission and intentionally omits any author and institution information.

The released Parquet contains 19,421 English examples.

Supported Tasks and Leaderboards

• Task: Three-way bias classification on policy and policy-adjacent excerpts
• Metrics: Accuracy, macro F1, precision, recall
• Evaluation views: Overall scores plus breakdowns by register, era, source, and bias_label
• Leaderboard: No public leaderboard is bundled in this repository; the dataset is intended for offline evaluation and analysis

Languages

• English

Dataset Structure

Data Fields

• sentence_id (string): Unique sequential identifier
• text (string): The excerpt to classify
• bias_label (string): Primary label, one of harmful, harmless, or antibias
• source (string): Canonical source family
• register (string): formal or informal
• era (string): Source-era bucket such as modern, contemporary, pre-1965, or 1930s-1970s
• domain (string): Main domain type label for the main bias text

Data Splits

This release is provided as a single benchmark set rather than a train/validation/test package.

High-Level Composition

Considerations for Using the Data

Limitations

• The dataset mixes legal, policy, government, and social-media style text, so performance may vary by genre.
• Labels were harmonized from multiple upstream sources, so some annotation noise and label drift are possible.
• The label distribution is not perfectly balanced.
• Some examples contain offensive or harmful language.
• The dataset is best used for evaluation and robustness analysis, not as a standalone measure of fairness.

Licensing Information

BRIDGE is a composite dataset built from multiple upstream sources, and therefore is released as a license-preserving aggregated benchmark. The benchmark contains source-derived text from datasets released under multiple licenses, including CC BY-SA 4.0, CC BY 4.0, CC BY-NC-SA 4.0, and CC0. We therefore do not apply a single overriding license to all source-derived text. Each example includes source attribution and source-level license metadata, and source-derived text remains governed by its original source license. Newly created BRIDGE annotations, labels, metadata, prompts, and evaluation code are released under CC BY-NC-SA 4.0 unless otherwise noted. Users are responsible for complying with the license associated with each example.

Citation

[CITATION WITHHELD FOR ANONYMOUS REVIEW]