{
 "body": "\n## What it measures\n\nHELM's decodingtrust_stereotype_bias run shows an English stereotype and asks the model to agree or disagree. DecodingTrust section 4 wrote 1,152 user prompts from 16 topics, three phrasings each, and 24 demographic groups. HELM expands that to 3,456 instances by repeating each prompt under benign, untargeted, and targeted system-prompt tags. Every item is a stereotype, so the intended behaviour is disagreement. This is not [bbq](bbq.md) ambiguous QA and not [crows_pairs](crows_pairs.md) likelihood comparison.\n\n## How it is scored\n\nThe paper's agreementIndex is agrees / 25 completions. HELM schema text treats `decodingtrust_stereotype_bias` as one minus that mean (higher is less agreement) plus a rejection rate. StereotypeMetric instead adds `1 \u2212 (agree_matrix.mean() / n_completions)` after every instance while a 24\u00d716\u00d73 count matrix fills, so the HELM Stat is a mean of running values, not one grid-wide 1\u2212agreementIndex. Completions are keyword-classified; mixed agree+disagree strings fall through to NO_RESPONSE except for a few \"as an AI\" corners. Temperature is 1 with 25 samples, matching the paper's n=25. Schema `main_split` is valid; the scenario emits TEST_SPLIT.\n\n## Dataset and licence\n\n1,152 unique user prompts; 3,456 HELM rows after three system types. Licence is CC-BY-SA-4.0. The jsonl is public on Data-Legacy; Hugging Face is gated. The benign tag in the file HELM downloads is spelled `beningn`. StereotypeMetric looks up `benign`, so a stock HELM score on that jsonl can raise ValueError unless the tag is patched.\n\n## Who publishes it\n\nDecodingTrust authors as on arXiv:2306.11698 (NeurIPS 2023 Datasets and Benchmarks Track). Prompts are original to this paper, not BBQ. HELM wrap: Stanford CRFM. Project site: decodingtrust.github.io.\n\n## Lineage\n\nSiblings with pages: [decodingtrust_adv_robustness](decodingtrust_adv_robustness.md), [decodingtrust_adv_demonstration](decodingtrust_adv_demonstration.md), [decodingtrust_fairness](decodingtrust_fairness.md), [decodingtrust_privacy](decodingtrust_privacy.md), [decodingtrust_machine_ethics](decodingtrust_machine_ethics.md), [decodingtrust_ood_robustness](decodingtrust_ood_robustness.md), [decodingtrust_toxicity_prompts](decodingtrust_toxicity_prompts.md). Appendix B.3 of the paper also discusses a standard stereotype benchmark; that extra protocol is not this HELM id.\n\n## Saturation and contamination\n\nUnknown current saturation. The statements have been public since 2023. Phrase-matching under-counts hedges that never say \"I agree\" verbatim. HELM entered maintenance mode on 2026-06-01.\n\n## How to run it\n\nHELM: `decodingtrust_stereotype_bias` (run entries add `task=all`, which the scenario ignores). Do not compare HELM's inverted 1\u2212agreementIndex to a paper heatmap of raw agreementIndex. Confirm whether the loader rewrites `beningn` before trusting a HELM number. lm-eval and inspect_evals names were not found.\n\n## Reading the numbers\n\nA high HELM bias score here means low keyword agreement with these 1,152 statements, not a general fairness certificate. Targeted system prompts are a different condition from benign. 25 samples at temperature 1 are part of the metric; a single greedy decode is not the same task. Read rejection rate beside the inverted agreement: a model that always refuses can look unbiased.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "decodingtrust_stereotype_bias",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [],
 "page": {
  "aliases": [
   "DecodingTrust - Stereotype Bias",
   "DecodingTrustStereotypeBiasScenario"
  ],
  "category": "safety",
  "contamination": {
   "note": "Manually written prompts, public in Data-Legacy since 2023, listed on gated Hugging Face AI-Secure/DecodingTrust. No measured memorisation study was found.\n",
   "risk": "medium"
  },
  "dataset": {
   "languages": [
    "en"
   ],
   "license": "CC-BY-SA-4.0",
   "modalities": [
    "text"
   ],
   "public_test_set": true,
   "size": 3456,
   "size_note": "Direct count of Data-Legacy stereotype_bias_data.jsonl at HELM pin 38972f6ccbf376a8d0660babafb4d2b3b9cca3f4: 3,456 lines = 1,152 user prompts \u00d7 3 system types. Paper and the stereotype README: 1,152 user prompts (16 topics \u00d7 3 variants \u00d7 24 groups). HELM tags copy sys_prompt_type_tag from the jsonl; the benign value is misspelled \"beningn\" in that file (1,152 rows).\n",
   "splits": "HELM TEST_SPLIT; one jsonl, three system-prompt tags",
   "url": "https://huggingface.co/datasets/AI-Secure/DecodingTrust"
  },
  "freshness": {
   "researched": "2026-09-08",
   "researched_by": "Grok Build, batch-038 (Codex coordinated)",
   "reviewed": "2026-09-08",
   "reviewed_by": "Grok Build independent review, batch-038"
  },
  "harness": {
   "bigbench": "",
   "helm": "decodingtrust_stereotype_bias",
   "inspect_evals": "",
   "lm_eval": "",
   "opencompass": "",
   "other": "Run spec decodingtrust_stereotype_bias (run entries pass task=all, ignored by the scenario). 25 samples per prompt, temperature 1, max_tokens 150.\n"
  },
  "id": "decodingtrust_stereotype_bias",
  "last_updated": "2024-02",
  "leaderboard_url": "https://decodingtrust.github.io/",
  "lineage": {
   "family": "",
   "predecessor": "",
   "successors": [],
   "variants": [
    "decodingtrust_adv_demonstration",
    "decodingtrust_adv_robustness",
    "decodingtrust_fairness",
    "decodingtrust_privacy"
   ]
  },
  "measures": "decodingtrust_stereotype_bias is HELM's wrap of DecodingTrust section 4. Each English user prompt states a stereotype and tells the model to append agree or disagree. There are 16 topics (HIV, terrorists, drug addiction, intelligence, greed, parenting, country, technology, weakness, driving, crime, drug dealing, jobs, leadership, STEM, hygiene) and 24 demographic groups spanning race/ethnicity, gender/orientation, nationality, age, religion, disability, and socioeconomic status. Three system-prompt types \u2014 benign, untargeted jailbreak, and targeted jailbreak \u2014 are stored as tags. Every statement is a stereotype; the paper treats agreement as the biased outcome.\n",
  "metric": {
   "baseline_note": "Paper section 4.2 defines agreementIndex = n_agree / 25 in [0, 1] (higher means more agreement with the stereotype). HELM schema_decodingtrust.yaml names decodingtrust_stereotype_bias as 1 \u2212 agreementIndex and marks it higher-is-better, plus decodingtrust_stereotype_bias_rejection_rate. StereotypeMetric classifies each completion with agree / disagree / no-response phrase lists, then after every instance adds 1 \u2212 (agree_matrix.mean() / n_completions) as a 24\u00d716\u00d73 count matrix fills, so the HELM Stat is a mean of running values, not one paper-style index. No random or human baseline is published.\n",
   "direction": "higher_is_better",
   "human_baseline": null,
   "max_score": 1.0,
   "name": "decodingtrust_stereotype_bias",
   "random_baseline": null,
   "unit": ""
  },
  "name": "DecodingTrust Stereotype Bias",
  "page_kind": "benchmark",
  "paper": {
   "arxiv": "2306.11698",
   "title": "DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models",
   "url": "https://arxiv.org/abs/2306.11698",
   "year": 2023
  },
  "publisher": {
   "authors": [
    "Boxin Wang",
    "Weixin Chen",
    "Hengzhi Pei",
    "Chulin Xie",
    "Mintong Kang",
    "Chenhui Zhang",
    "Chejian Xu",
    "Zidi Xiong",
    "Ritik Dutta",
    "Rylan Schaeffer",
    "Sang T. Truong",
    "Simran Arora",
    "Mantas Mazeika",
    "Dan Hendrycks",
    "Zinan Lin",
    "Yu Cheng",
    "Sanmi Koyejo",
    "Dawn Song",
    "Bo Li"
   ],
   "org": "DecodingTrust authors (UIUC / Stanford / collaborators); HELM wrap by Stanford CRFM",
   "url": "https://decodingtrust.github.io/"
  },
  "released": "2023-06",
  "repo_url": "https://github.com/AI-secure/DecodingTrust",
  "saturation": {
   "as_of": "",
   "note": "Paper section 4 reports 2023 GPT-3.5/GPT-4 agreement heatmaps, not a single ceiling. No current HELM DecodingTrust leaderboard URL resolved (404 on crfm.stanford.edu/helm/decodingtrust/latest/, 2026-09-08). HELM README states maintenance mode from 2026-06-01. StereotypeMetric indexes SYS_PROMPT_TYPE [\"benign\",\"untargeted\",\"targeted\"]; the jsonl's \"beningn\" tag is not in that list.\n",
   "status": "unknown",
   "top_score": null
  },
  "sources": [
   {
    "accessed": "2026-09-08",
    "title": "HELM DecodingTrustStereotypeBiasScenario (1,152-prompt docstring; jsonl tags)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/scenarios/decodingtrust_stereotype_bias_scenario.py"
   },
   {
    "accessed": "2026-09-08",
    "title": "HELM StereotypeMetric (agreement phrases; 1 \u2212 mean agree)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/metrics/decodingtrust_stereotype_bias_metrics.py"
   },
   {
    "accessed": "2026-09-08",
    "title": "HELM decodingtrust_stereotype_bias run spec (25 outputs, temperature 1)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/run_specs/decodingtrust_run_specs.py"
   },
   {
    "accessed": "2026-09-08",
    "title": "HELM schema_decodingtrust.yaml (1 \u2212 agreementIndex)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/static/schema_decodingtrust.yaml"
   },
   {
    "accessed": "2026-09-08",
    "title": "Data-Legacy stereotype README (1,152 user prompts, 25 generations)",
    "url": "https://raw.githubusercontent.com/AI-secure/DecodingTrust-Data-Legacy/38972f6ccbf376a8d0660babafb4d2b3b9cca3f4/data/stereotype/README.md"
   },
   {
    "accessed": "2026-09-08",
    "title": "DecodingTrust HTML section 4.2 (agreementIndex, n=25)",
    "url": "https://ar5iv.labs.arxiv.org/html/2306.11698"
   },
   {
    "accessed": "2026-09-08",
    "title": "DecodingTrust paper (arXiv:2306.11698)",
    "url": "https://arxiv.org/abs/2306.11698"
   },
   {
    "accessed": "2026-09-08",
    "title": "Hugging Face DecodingTrust README (stereotype config, CC-BY-SA-4.0)",
    "url": "https://huggingface.co/datasets/AI-Secure/DecodingTrust/resolve/main/README.md"
   },
   {
    "accessed": "2026-09-08",
    "title": "DecodingTrust project site",
    "url": "https://decodingtrust.github.io/"
   },
   {
    "accessed": "2026-09-08",
    "title": "HELM README (maintenance mode from 2026-06-01)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/README.md"
   }
  ],
  "status": "unknown",
  "subcategory": "HELM wrap of DecodingTrust section 4 (1,152 stereotype statements \u00d7 3 system prompts)",
  "summary": "HELM scenario for DecodingTrust stereotypes: agree/disagree with 1,152 statements across 16 topics, 24 groups, and 3 system prompts.",
  "tags": [
   "safety",
   "bias",
   "stereotype",
   "helm",
   "decodingtrust"
  ],
  "task_format": "HELM instruct generation, num_outputs 25, max_tokens 150, temperature 1. The scenario loads stereotype_bias_data.jsonl and does not take a task argument (the run function still declares unused task: str).\n"
 }
}