{
 "body": "\n## What it measures\n\nThis id is HELM Instruct's use of Anthropic's HH-RLHF dialogues, not Anthropic's training recipe. The model is given the first Human: turn only and must answer it. HELM strips later turns and ignores which completion was chosen. Subset `hh` draws those first turns from the public preference files. Subset `red_team` draws them from red-team transcripts whose original train split HELM treats as test. The language is English. The intended skill is following that request, including when the request is a harmful probe.\n\n## How it is scored\n\n`InstructionFollowingCritiqueMetric` collects human (or HELM critique-backend) ratings on five 1\u20135 axes. Helpfulness is the schema headline: whether the model appears to do what it was instructed to do. Harmlessness is a separate axis and can pull the other way on red-team prompts. Official `run_entries.conf` sets `num_respondents=1`. Empty references mean there is no exact-match to a gold assistant. Bai et al. and Ganguli et al. report preference-model and red-team analyses, not this 1\u20135 critique; do not mix those figures with HELM Helpfulness.\n\n## Dataset and licence\n\nAnthropic/hh-rlhf is MIT-licensed. The default preference config has 160,800 train and 8,552 test pairs on datasets-server. HELM `hh` walks every pair and keeps the first utterance of `chosen`. Ganguli et al. state 38,961 red-team attacks; HELM `red_team` loads that directory and scores it all as test. Unique prompts may be fewer than pair counts. The HF card warns that the data contain offensive content and are not meant for supervised training of dialogue agents.\n\n## Who publishes it\n\nPreference data: Bai, Jones, Ndousse, Askell, and co-authors, \"Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback\" (arXiv:2204.05862, April 2022). Red-team data: Ganguli, Lovitt, Kernion, and co-authors (arXiv:2209.07858, September 2022). Hugging Face copy created 8 December 2022. HELM packaging is Stanford CRFM's Instruct leaderboard (`https://crfm.stanford.edu/helm/instruct/`). The original GitHub tree is deprecated in favour of the HF dataset.\n\n## Lineage\n\nHELM Instruct groups this scenario with [self_instruct](self_instruct.md), Vicuna, Koala, Open Assistant, and grammar, all sharing the same critique metric. It is not [helm_safety](helm_safety.md): that leaderboard's `anthropic_red_team` scenario scores the same red-team first turns with `safety_score`, not this Helpfulness critique. Using HH-RLHF to train a reward model is a different job from this eval.\n\n## Saturation and contamination\n\nNo HELM Instruct top cell was read from the JavaScript leaderboard. Contamination risk is high. The prompts have been public since 2022 and are standard preference-tuning data. A model that was RLHF'd on HH-RLHF may look strong here because it has seen the first turns, not because the critique protocol is hard.\n\n## How to run it\n\nHELM run spec `anthropic_hh_rlhf:subset=hh` or `subset=red_team`, with `num_respondents` and a working critique backend. Generation is 512 tokens at temperature 0.7, zero-shot. Without enough critiques the metric returns nothing, so a missing Helpfulness cell is not a zero. HELM entered maintenance mode on 1 June 2026. lm-evaluation-harness and Inspect Evals were not confirmed to ship this HELM critique task under this id.\n\n## Reading the numbers\n\nA high Helpfulness mean means the critique rater thought the model followed the first human turn, on a 1\u20135 rubric, with the configured respondent count. It is not win-rate against `rejected`, not harmlessness, and not HELM Safety. On `red_team`, a helpful answer to a harmful request can raise Helpfulness while lowering Harmlessness; read both axes. Compare only the same subset, respondent count, and temperature.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "anthropic_hh_rlhf",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [],
 "page": {
  "aliases": [
   "HH-RLHF",
   "Anthropic RLHF dataset",
   "anthropic_hh_rlhf:subset=hh",
   "anthropic_hh_rlhf:subset=red_team"
  ],
  "category": "instruction-following",
  "contamination": {
   "note": "Prompts have been public since 2022 (HF dataset created 8 December 2022; papers April and September 2022). The files are widely reused in preference-tuning. HELM Instruct scores the first human turn of those public transcripts. HELM Safety ships a separate anthropic_red_team scenario on the same red-team transcripts, scored with safety_score rather than this Helpfulness critique. Ganguli et al. release 38,961 attacks.\n",
   "risk": "high"
  },
  "dataset": {
   "languages": [
    "en"
   ],
   "license": "MIT",
   "modalities": [
    "text"
   ],
   "public_test_set": true,
   "size": null,
   "size_note": "HELM `subset=hh` loads Anthropic/hh-rlhf default (revision 09be8c5bbc57cb3887f3a9732ad6aa7ec602a1fa) and keeps the first utterance of each `chosen` field. Hugging Face datasets-server reports 160,800 train and 8,552 test rows (169,352 total) for that default preference config. HELM `subset=red_team` loads data_dir red-team-attempts and maps the original train split to HELM test. Ganguli et al. (arXiv 2209.07858 abstract) release 38,961 red-team attacks; datasets-server size for that config returned HTTP 500 when requested. Unique first utterances may be fewer than row counts because many pairs share a prompt. HELM does not deduplicate in the scenario source.\n",
   "splits": "hh: HF train and test kept; red_team: original train used as HELM test",
   "url": "https://huggingface.co/datasets/Anthropic/hh-rlhf"
  },
  "freshness": {
   "researched": "2026-09-08",
   "researched_by": "Grok Build, batch-025 (Codex coordinated)",
   "reviewed": "2026-09-08",
   "reviewed_by": "Grok Build independent review, batch-025"
  },
  "harness": {
   "bigbench": "",
   "helm": "anthropic_hh_rlhf",
   "inspect_evals": "",
   "lm_eval": "",
   "opencompass": "",
   "other": "Run spec anthropic_hh_rlhf:subset=hh|red_team in instruction_following_run_specs.py; classic run_entries.conf uses num_respondents=1 (hh priority 1, red_team priority 3)."
  },
  "id": "anthropic_hh_rlhf",
  "last_updated": "2023-05",
  "leaderboard_url": "https://crfm.stanford.edu/helm/instruct/latest/",
  "lineage": {
   "family": "",
   "predecessor": "",
   "successors": [],
   "variants": []
  },
  "measures": "anthropic_hh_rlhf, as this id, is Stanford CRFM's HELM Instruct scenario over Anthropic's public HH-RLHF dialogues. The model sees only the first human utterance of a conversation and must write a free-form English reply. HELM does not present the assistant turns, does not use chosen versus rejected labels as the score, and does not run RLHF. Two subsets exist: `hh` (helpfulness/harmlessness preference dialogues) and `red_team` (red-team transcripts). The skill is following that first request, under a human critique, not preference-model accuracy.\n",
  "metric": {
   "baseline_note": "Annotators also rate Understandability, Completeness, Conciseness, and Harmlessness on the same 1-5 maps. schema_instruction_following.yaml sets main_name Helpfulness (\"Does the model appear to do what it is instructed to?\"). If fewer than num_respondents critiques return, the metric emits no stats. This is not reward-model accuracy on chosen versus rejected completions.\n",
   "direction": "higher_is_better",
   "human_baseline": null,
   "max_score": 5.0,
   "name": "Helpfulness",
   "random_baseline": null,
   "unit": "1-5"
  },
  "name": "Anthropic HH-RLHF (HELM Instruct)",
  "page_kind": "benchmark",
  "paper": {
   "arxiv": "2204.05862",
   "title": "Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback",
   "url": "https://arxiv.org/abs/2204.05862",
   "year": 2022
  },
  "publisher": {
   "authors": [
    "Yuntao Bai",
    "Andy Jones",
    "Kamal Ndousse",
    "Amanda Askell",
    "Deep Ganguli"
   ],
   "org": "Anthropic (dataset); Stanford CRFM (HELM Instruct scenario)",
   "url": "https://huggingface.co/datasets/Anthropic/hh-rlhf"
  },
  "released": "2022-04",
  "repo_url": "https://github.com/stanford-crfm/helm/blob/main/src/helm/benchmark/scenarios/anthropic_hh_rlhf_scenario.py",
  "saturation": {
   "as_of": "",
   "note": "No numeric HELM Instruct cell was read (the public page is a JavaScript app). Anthropic's papers report RLHF training results, not this critique protocol.\n",
   "status": "unknown",
   "top_score": null
  },
  "sources": [
   {
    "accessed": "2026-09-08",
    "title": "HELM AnthropicHHRLHFScenario (first utterance only; hh vs red_team)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/scenarios/anthropic_hh_rlhf_scenario.py"
   },
   {
    "accessed": "2026-09-08",
    "title": "HELM anthropic_hh_rlhf run spec (Instruct adapter, critique metrics)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/run_specs/instruction_following_run_specs.py"
   },
   {
    "accessed": "2026-09-08",
    "title": "InstructionFollowingCritiqueMetric (1-5 Helpfulness and four other axes)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/metrics/instruction_following_critique_metrics.py"
   },
   {
    "accessed": "2026-09-08",
    "title": "HELM Instruct schema (main_name Helpfulness; anthropic_hh_rlhf group)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/static/schema_instruction_following.yaml"
   },
   {
    "accessed": "2026-09-08",
    "title": "HELM classic run entries (hh and red_team, num_respondents=1)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/presentation/run_entries.conf"
   },
   {
    "accessed": "2026-09-08",
    "title": "get_instruct_adapter_spec (512 tokens, temperature 0.7, zero-shot)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/adaptation/common_adapter_specs.py"
   },
   {
    "accessed": "2026-09-08",
    "title": "Anthropic/hh-rlhf dataset card (MIT; PM data vs red-team data)",
    "url": "https://huggingface.co/datasets/Anthropic/hh-rlhf/raw/main/README.md"
   },
   {
    "accessed": "2026-09-08",
    "title": "HF API (licence mit, created 2022-12-08, revision 09be8c5b\u2026)",
    "url": "https://huggingface.co/api/datasets/Anthropic/hh-rlhf"
   },
   {
    "accessed": "2026-09-08",
    "title": "datasets-server default config size (160800 train / 8552 test)",
    "url": "https://datasets-server.huggingface.co/size?dataset=Anthropic/hh-rlhf"
   },
   {
    "accessed": "2026-09-08",
    "title": "Bai et al. HH-RLHF paper (arXiv:2204.05862)",
    "url": "https://arxiv.org/abs/2204.05862"
   },
   {
    "accessed": "2026-09-08",
    "title": "Ganguli et al. red-teaming paper (38,961 attacks; arXiv:2209.07858)",
    "url": "https://arxiv.org/abs/2209.07858"
   },
   {
    "accessed": "2026-09-08",
    "title": "anthropics/hh-rlhf GitHub README (deprecated in favour of the HF copy)",
    "url": "https://raw.githubusercontent.com/anthropics/hh-rlhf/master/README.md"
   },
   {
    "accessed": "2026-09-08",
    "title": "HELM AnthropicRedTeamScenario (same red-team transcripts; safety_score)",
    "url": "https://raw.githubusercontent.com/stanford-crfm/helm/main/src/helm/benchmark/scenarios/anthropic_red_team_scenario.py"
   },
   {
    "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 Instruct critique of first-turn prompts from Anthropic HH-RLHF",
  "summary": "HELM Instruct scores first human utterances from Anthropic HH-RLHF with a 1-5 Helpfulness critique; it does not train on or rank the chosen/rejected pairs.",
  "tags": [
   "instruction-following",
   "helm",
   "critique",
   "hh-rlhf",
   "red-teaming"
  ],
  "task_format": "Zero-shot generation via get_instruct_adapter_spec (max_tokens 512, temperature 0.7, max_train_instances 0). References are empty. Scoring is InstructionFollowingCritiqueMetric with caller-chosen num_respondents. Official classic run entries use num_respondents=1.\n"
 }
}