{
 "body": "\nPart of the [ARC](arc.md) family, alongside its sibling split [ARC-Challenge](arc_challenge.md).\n\n## What it measures\n\nARC-Easy is the larger, easier half of the AI2 Reasoning Challenge: 5,197 grade-school-level natural\nscience exam questions that at least one of two 2018-era baseline solvers (information retrieval or\nword-co-occurrence/PMI) could already answer correctly. Everything both baselines failed went to\nARC-Challenge instead. Format, authorship and licensing are identical to Challenge -- only the difficulty\nfilter differs -- so a bare \"ARC\" score with no split named should not be trusted until you confirm which\nhalf it comes from.\n\n## Reading the numbers\n\nARC-Easy is now saturated for modern models: open-weight base models from late 2024 already score in\nthe mid-90s to 98% under a 5-shot, character-normalised protocol on a 1,000-item test subsample,\nessentially at ceiling given the split's own construction (it only contains questions weak baselines\ncould already solve). A high ARC-Easy score today confirms a model is not badly broken on basic science\nQA, but it separates almost no current models from each other and should never substitute for an\nARC-Challenge, MMLU-Pro or GPQA-Diamond score as evidence of reasoning ability.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "arc_easy",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [],
 "page": {
  "aliases": [
   "ARC (Easy Set)",
   "AI2 Reasoning Challenge (Easy)"
  ],
  "category": "reasoning",
  "contamination": {
   "note": "Questions and answers have been fully public since 2018 and are widely mirrored, so any model trained on a broad web crawl since then has plausibly seen them; no dedicated post-hoc contamination study specific to this split was found.",
   "risk": "medium"
  },
  "dataset": {
   "languages": [
    "en"
   ],
   "license": "CC-BY-SA-4.0",
   "modalities": [
    "text"
   ],
   "public_test_set": true,
   "size": 5197,
   "size_note": "5,197 questions (2,251 train / 570 dev / 2,376 test), out of 7,787 total ARC questions.",
   "splits": "train (2,251) / dev (570) / test (2,376)",
   "url": "https://huggingface.co/datasets/allenai/ai2_arc"
  },
  "freshness": {
   "researched": "2026-09-08",
   "researched_by": "sonnet-5 agent, batch 4, slice D",
   "reviewed": "",
   "reviewed_by": ""
  },
  "harness": {
   "bigbench": "",
   "helm": "",
   "inspect_evals": "arc_easy",
   "lm_eval": "arc_easy",
   "opencompass": "ARC-e",
   "other": ""
  },
  "id": "arc_easy",
  "last_updated": "",
  "leaderboard_url": "",
  "lineage": {
   "family": "arc",
   "predecessor": "",
   "successors": [],
   "variants": []
  },
  "measures": "ARC-Easy is the larger of the two AI2 Reasoning Challenge splits: 5,197 grade-school-level natural science exam questions that landed here specifically because at least one of two 2018-era baseline solvers (an information-retrieval solver or a word-co-occurrence/PMI solver) answered them correctly. Every question that both baselines failed went to the harder ARC-Challenge split instead. Format, source and authorship are identical to Challenge; only the difficulty filter differs.\n",
  "metric": {
   "baseline_note": "By construction, contemporary (2018) retrieval and PMI baselines could already solve this split, unlike Challenge, so it was never intended as a difficult bar for modern models.",
   "direction": "higher_is_better",
   "human_baseline": null,
   "max_score": 100.0,
   "name": "accuracy (often reported as acc_norm, length-normalised)",
   "random_baseline": 25.0,
   "unit": "%"
  },
  "name": "ARC-Easy",
  "page_kind": "subset",
  "paper": {
   "arxiv": "1803.05457",
   "title": "Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge",
   "url": "https://arxiv.org/abs/1803.05457",
   "year": 2018
  },
  "publisher": {
   "authors": [
    "Peter Clark",
    "Isaac Cowhey",
    "Oren Etzioni",
    "Tushar Khot",
    "Ashish Sabharwal",
    "Carissa Schoenick",
    "Oyvind Tafjord"
   ],
   "org": "Allen Institute for AI (AI2)",
   "url": "https://allenai.org/"
  },
  "released": "2018-03",
  "repo_url": "https://huggingface.co/datasets/allenai/ai2_arc",
  "saturation": {
   "as_of": "2024-12",
   "note": "The OLMo 2 technical report (arXiv:2501.00656, posted 2024-12-31) reports several open-weight base models already scoring in the low-to-high 90s on a 1,000-item subsample of the 2,376-item test set under the OLMES 5-shot, character-normalised protocol: Qwen2.5-14B 98.2%, Zamba-2-7B 96.7%, Qwen2.5-7B 96.1%, Gemma-2-9B 95.5%. This is a different protocol (subsampled, char-normalised, 5-shot) from lm-evaluation-harness's default acc_norm on the full test set, so read it as strong directional evidence of saturation rather than a directly comparable ceiling figure.",
   "status": "saturated",
   "top_score": 98.2
  },
  "sources": [
   {
    "accessed": "2026-09-08",
    "title": "Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge",
    "url": "https://arxiv.org/abs/1803.05457"
   },
   {
    "accessed": "2026-09-08",
    "title": "allenai/ai2_arc dataset metadata (Hugging Face API)",
    "url": "https://huggingface.co/api/datasets/allenai/ai2_arc"
   },
   {
    "accessed": "2026-09-08",
    "title": "lm-evaluation-harness arc task directory",
    "url": "https://github.com/EleutherAI/lm-evaluation-harness/tree/main/lm_eval/tasks/arc"
   },
   {
    "accessed": "2026-09-08",
    "title": "OpenCompass ARC_e config directory (abbr `ARC-e`, loads opencompass/ai2_arc-easy-dev)",
    "url": "https://github.com/open-compass/opencompass/tree/main/opencompass/configs/datasets/ARC_e"
   },
   {
    "accessed": "2026-09-08",
    "title": "\"2 OLMo 2 Furious\" technical report (ar5iv) -- ARC-Easy scores for contemporary open-weight base models under the OLMES protocol",
    "url": "https://ar5iv.labs.arxiv.org/html/2501.00656"
   }
  ],
  "status": "superseded",
  "subcategory": "grade-school science multiple-choice QA (Easy split)",
  "summary": "The easier, larger half of the AI2 Reasoning Challenge; questions that 2018-era retrieval or word-overlap baselines could already answer, now scored near ceiling by current models.",
  "tags": [
   "science-qa",
   "multiple-choice",
   "reasoning",
   "subset",
   "legacy-benchmark",
   "saturated"
  ],
  "task_format": "Multiple-choice science question, typically 4 answer options, single correct answer."
 }
}