dataset_info:
features:

• name: test_date
  dtype: large_string
• name: model
  dtype: large_string
• name: prompt_id
  dtype: large_string
• name: category
  dtype: large_string
• name: style
  dtype: large_string
• name: persona
  dtype: large_string
• name: verdict
  dtype: large_string
• name: classification
  dtype: large_string
• name: prompt_text
  dtype: large_string
• name: response_text
  dtype: large_string
• name: prompt_tokens
  dtype: large_string
• name: completion_tokens
  dtype: large_string
• name: total_tokens
  dtype: large_string
• name: run_cost
  dtype: float64
• name: confidence
  dtype: float64
• name: reasoning
  dtype: large_string
  splits:
• name: train
  num_bytes: 249148674
  num_examples: 294519
  download_size: 35073976
  dataset_size: 249148674
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
    annotations_creators:
• machine-generated
  license: mit
  source_datasets:
• original
  tags:
• ai-safety
• bias-analysis
• censorship
• content-moderation
• inter-annotator-agreement
• llm
• over-refusal
• red-teaming
  pretty_name: 'Moderation Bias: LLM Content Moderation Benchmark'
  size_categories:
• 100K<n<1M
  task_categories:
• text-classification
  language:
• en

Moderation Bias: LLM Content Moderation Benchmark

One row per model evaluation of one prompt. This dataset is the raw audit log
behind moderationbias.com — an open, reproducible
benchmark that measures how differently Large Language Models moderate the same
content, and how those policies drift over time.

• Homepage: https://moderationbias.com
• Repository: https://github.com/jacobkandel/llm-content-moderation-analysis
• Leaderboard: https://moderationbias.com/leaderboard
• Paper / Methodology: https://moderationbias.com/methodology · METHODOLOGY.md
• Datasheet: DATASHEET.md
• Point of contact: Jacob Kandel (ORCID 0009-0008-8858-6072)
• DOI: 10.5281/zenodo.20262255
• License: MIT

Dataset Summary

Moderation Bias audits content-moderation behavior across 30 Large Language Models,
16 harm categories, and 2,012 evaluated prompts (2,323 in the corpus), with
biweekly automated re-runs that track policy drift. Each prompt is sent to every subject
model and its response is classified by a held-out LLM judge into a canonical verdict
(ALLOWED vs. a family of refusal labels). The result is a large, longitudinal panel of
moderation decisions suitable for measuring inter-model disagreement, over-refusal,
category-level bias, and drift.

Headline finding: model identity (Cramér's V = 0.66) is a substantially stronger
predictor of refusal than prompt topic (V = 0.20) — evidence of provider policy
idiosyncrasy rather than rational, content-based moderation.

Supported Tasks

• text-classification — predict the moderation verdict (ALLOWED / refusal) for a
  prompt–response pair, or the finer classification label
  (Authorized / Hard_Refusal / Soft_Censorship / False_Refusal).
• Model auditing / evaluation — compare refusal rates and disagreement across models
  and categories; measure over-refusal against safe-prompt benchmarks (XSTest, OR-Bench).
• Drift analysis — track how a given model's verdicts change across audit dates.

Languages

Primarily English (en). A dedicated INTL category and translated variants probe
cross-lingual moderation; those rows carry non-English prompt text.

Dataset Structure

Data Fields

One row per (model, prompt, run). Text columns are present in the full export and empty
in the lite export.

A refusal rate is # refusals / # scorable rows; ERROR rows are excluded from
the denominator. Repetitions/variants of a seed prompt are correlated — pairwise tests
collapse them to one majority verdict per (prompt, model).

Splits

The dataset is published as a single consolidated table (no train/test split); it is an
audit log, not a supervised training set. Users constructing a classification task should
create their own splits and be mindful that variants of the same seed prompt are
correlated (split by base prompt_id, not by row).

Dataset Creation

Curation Rationale

Public moderation benchmarks tend to measure whether a model refuses unsafe content.
This benchmark instead measures disagreement and over-refusal: where models diverge on
identical prompts, and where safe prompts are wrongly refused. Prompts are stratified
across 16 harm categories and paired with safe-prompt controls to separate genuine safety
behavior from paternalistic over-restriction.

Source Data

Prompts combine an internally authored corpus with items drawn from established
safe-prompt / over-refusal benchmarks (e.g. XSTest, OR-Bench); benchmark_source
records each prompt's origin. Responses are generated live via the OpenRouter API across
the subject-model registry.

Annotations (verdicts)

Verdicts are produced by an LLM-as-judge that is held out of the subject pool to
avoid self-preference bias (see METHODOLOGY §2.2). Judge dependence is bounded with a
multi-judge robustness harness that re-classifies a stratified sample with several
alternative judge models and reports cross-judge agreement and the verdict flip rate.
Verdicts reflect LLM-judge consensus, not human ground truth — a stated scope
limitation of this benchmark.

Personal and Sensitive Information

Prompts are synthetic or drawn from public benchmarks and are not about real
individuals. Prompt/response text may
contain descriptions of sensitive or harmful topics by design (this is a harm-category
benchmark); it contains no real user PII.

Considerations for Using the Data

Social Impact & Intended Use

Intended for research and auditing of moderation policy — transparency into how
commercial models restrict content. It is not a training target for making models more
or less restrictive, and refusal ≠ "correct" moderation.

Known Limitations & Biases

• Judge dependence. Verdicts reflect a single LLM judge's labeling; the human-alignment
  and multi-judge robustness checks quantify but do not eliminate this. See
  LIMITATIONS.md.
• Corpus scope. English-dominant; the harm taxonomy and prompt selection encode the
  authors' framing of "moderation-relevant" content.
• Correlated rows. Variants/repetitions of a seed prompt are not independent.
• Provider/version drift. model/model_version behavior changes over time; always
  condition on test_date for longitudinal claims.

Additional Information

Licensing

Released under the MIT License.

Citation

@software{kandel_moderation_bias,
  author  = {Kandel, Jacob},
  title   = {Moderation Bias: LLM Content Moderation Benchmark},
  url     = {https://moderationbias.com},
  doi     = {10.5281/zenodo.20262255},
  version = {2.1.0},
  year    = {2026}
}

Contributions

Issues and contributions welcome at the
GitHub repository.
This card is generated from the repo's datasheet and data dictionary; the dataset itself
is refreshed by the biweekly audit pipeline.