{
  "benchmark_card": {
    "benchmark_details": {
      "name": "CivilComments",
      "overview": "CivilComments is a benchmark designed to measure unintended identity-based bias in toxicity classification models. It uses a large, real-world dataset of online comments from the Civil Comments platform, extended with crowd-sourced annotations for toxicity and demographic identity references. This provides a nuanced evaluation of bias beyond synthetic datasets.",
      "data_type": "tabular, text",
      "domains": [
        "machine learning fairness",
        "bias measurement",
        "toxic comment classification",
        "text classification"
      ],
      "languages": [
        "English"
      ],
      "similar_benchmarks": "The paper does not name other specific benchmark datasets, only referencing prior work using synthetic test sets.",
      "resources": [
        "https://arxiv.org/abs/1903.04561",
        "https://huggingface.co/datasets/google/civil_comments",
        "https://storage.googleapis.com/crfm-helm-public/benchmark_output/releases/v0.4.0/groups/core_scenarios.json"
      ]
    },
    "purpose_and_intended_users": {
      "goal": "To evaluate unintended identity-based bias in toxicity classification models using real data and nuanced metrics.",
      "audience": [
        "Machine learning researchers and practitioners working on fairness, bias measurement, and mitigation, particularly in toxic comment classification."
      ],
      "tasks": [
        "Binary toxicity classification (toxic vs. non-toxic)",
        "Analysis of performance across identity subgroups"
      ],
      "limitations": "The labeled set of identities is not comprehensive and does not provide universal coverage, representing a balance between coverage, annotator accuracy, and example count. The real-world data is potentially noisier than synthetic alternatives.",
      "out_of_scope_uses": [
        "Developing effective strategies for choosing optimal thresholds to minimize bias"
      ]
    },
    "data": {
      "source": "The data consists of online comments sourced from the Civil Comments platform, a commenting plugin for independent English-language news sites. The comments were publicly posted between 2015 and 2017.",
      "size": "The dataset contains approximately 1.8 million comments for training, with separate validation and test sets of approximately 97,320 examples each. All comments were labeled for toxicity, and a subset of 450,000 comments was additionally labeled for identity references.",
      "format": "parquet",
      "annotation": "Labeling was performed by crowd raters. Toxicity labels were applied using guidelines consistent with the Perspective API. For the identity-labeled subset, raters were shown comments and selected referenced identities (e.g., genders, races, ethnicities) from a provided list. Some comments for identity labeling were pre-selected by models to increase the frequency of identity content."
    },
    "methodology": {
      "methods": [
        "Models are evaluated by applying a suite of bias metrics to their predictions on the test set. The original paper demonstrates this using publicly accessible toxicity classifiers on the dataset."
      ],
      "metrics": [
        "Subgroup AUC",
        "BPSN AUC",
        "BNSP AUC",
        "Negative Average Equality Gap (AEG)",
        "Positive Average Equality Gap (AEG)"
      ],
      "calculation": "The evaluation calculates five metrics for each identity subgroup to provide a multi-faceted view of bias. There is no single aggregated overall score.",
      "interpretation": "For the AUC metrics (Subgroup, BPSN, BNSP), higher values indicate better separability (fewer mis-orderings). For the Average Equality Gaps (Negative and Positive), lower values indicate better separability (more similar score distributions).",
      "baseline_results": "Paper baselines: Results for TOXICITY@1 and TOXICITY@6 from the Perspective API are reported, showing their Subgroup AUC, BPSN AUC, BNSP AUC, Negative AEG, and Positive AEG on a synthetic dataset for the lowest performing 20 subgroups. They are also compared on short comments within the human-labeled dataset for specific identities. EEE results: Anthropic-LM v4-s3 52B scored 0.6100 on the CivilComments metric.",
      "validation": "The evaluation assumes the human-provided labels are reliable. The identity labeling set was designed to balance coverage, crowd rater accuracy, and ensure sufficient examples per identity for meaningful results."
    },
    "ethical_and_legal_considerations": {
      "privacy_and_anonymity": "The paper does not discuss how personally identifiable information (PII) in the online comments was handled or if data was anonymized.",
      "data_licensing": "Creative Commons Zero v1.0 Universal",
      "consent_procedures": "The paper does not describe compensation for crowdworkers or the specific platform used for annotation.",
      "compliance_with_regulations": "The paper does not mention IRB approval, GDPR compliance, or any other ethical review process."
    },
    "possible_risks": [
      {
        "category": "Unrepresentative data",
        "description": [
          "Unrepresentative data occurs when the training or fine-tuning data is not sufficiently representative of the underlying population or does not measure the phenomenon of interest. Synthetic data might not fully capture the complexity and nuances of real-world data. Causes include possible limitations in the seed data quality, biases in generation methods, or inadequate domain knowledge. Thus, AI models might struggle to generalize effectively to real-world scenarios."
        ],
        "url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/unrepresentative-data.html"
      },
      {
        "category": "Uncertain data provenance",
        "description": [
          "Data provenance refers to the traceability of data (including synthetic data), which includes its ownership, origin, transformations, and generation. Proving that the data is the same as the original source with correct usage terms is difficult without standardized methods for verifying data sources or generation."
        ],
        "url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/data-provenance.html"
      },
      {
        "category": "Data bias",
        "description": [
          "Historical and societal biases might be present in data that are used to train and fine-tune models. Biases can also be inherited from seed data or exacerbated by synthetic data generation methods."
        ],
        "url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/data-bias.html"
      },
      {
        "category": "Lack of data transparency",
        "description": [
          "Lack of data transparency might be due to insufficient documentation of training or tuning dataset details, including synthetic data generation.\u00a0"
        ],
        "url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/lack-of-data-transparency.html"
      },
      {
        "category": "Output bias",
        "description": [
          "Generated content might unfairly represent certain groups or individuals."
        ],
        "url": "https://www.ibm.com/docs/en/watsonx/saas?topic=SSYOK8/wsj/ai-risk-atlas/output-bias.html"
      }
    ],
    "flagged_fields": {},
    "missing_fields": [],
    "card_info": {
      "created_at": "2026-03-17T12:38:43.250822",
      "llm": "deepseek-ai/DeepSeek-V3.2"
    }
  }
}