{
 "body": "\nPart of the [CLUE](clue.md) family.\n\n## What it measures\n\nDRCD gives a model a Traditional Chinese Wikipedia paragraph and a question, requiring an extracted\nanswer span -- the same SQuAD-style format as [CMRC2018](clue_cmrc.md), but in Traditional rather\nthan Simplified characters and drawn from an entirely separate dataset. CLUE adopted DRCD unchanged\nfrom the Delta Reading Comprehension Dataset, published by a Delta Research Center team with no\nauthor overlap with the CLUE team -- the cleanest external adoption in this batch, unlike CMRC2018\nand C3 where the original authors also co-authored CLUE's paper. CLUE supplies its own\ntraditional-to-simplified conversion tool for models evaluated in simplified Chinese. DRCD is not\none of the paper's nine core tasks and is not a scored column on CLUE's live leaderboard; it\ncirculates mainly through CLUE's downloads and OpenCompass, not a maintained ranking.\n\n## Reading the numbers\n\nThe data holds 26,936 training, 3,524 development and 3,493 test questions (about 34,000 total,\nclose to the original paper's \"30,000+\" claim), confirmed identically by the CLUE README and the\nHugging Face mirror. Test answers are withheld (dummy text in the public mirror). DRCD's own paper\nreports a baseline F1 of 89.59% against 93.30% human F1 -- figures from the original dataset's own\nevaluation, not a CLUE leaderboard number, since CLUE does not score this task live. Compare a DRCD\nscore only against other DRCD scores, not CMRC2018's, despite the shared format.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "clue_drcd",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [],
 "page": {
  "category": "composite",
  "dataset": {
   "languages": [
    "zh"
   ],
   "license": "",
   "modalities": [
    "text"
   ],
   "public_test_set": false,
   "size": 33953,
   "size_note": "26,936 training, 3,524 development and 3,493 test questions (8,016 / 1,000 / 1,000 paragraphs respectively), confirmed identically by the CLUE GitHub README and the Hugging Face clue/clue mirror (config drcd). The original paper's abstract instead rounds to \"10,014 paragraphs...and 30,000+ questions\"; this page's paragraph total (10,016) is close but not identical, a small gap left unreconciled.\n",
   "splits": "train (26,936) / validation (3,524) / test (3,493, answers withheld in the public mirror)",
   "url": "https://huggingface.co/datasets/clue/clue"
  },
  "freshness": {
   "researched": "2026-09-08",
   "researched_by": "sonnet-5 agent, batch 5, slice A"
  },
  "harness": {
   "opencompass": "CLUE_DRCD (CLUE_DRCD_gen config family, several hash-suffixed revisions; loads the opencompass/drcd_dev mirror)"
  },
  "id": "clue_drcd",
  "leaderboard_url": "",
  "lineage": {
   "family": "clue"
  },
  "measures": "DRCD gives a model a Traditional Chinese Wikipedia paragraph and a question, requiring an extracted answer span -- the same SQuAD-style span-extraction format as CMRC2018, but in Traditional rather than Simplified characters, and drawn from an entirely separate dataset. CLUE adopted DRCD unchanged from the Delta Reading Comprehension Dataset, published by a Delta Research Center team with no author overlap with the CLUE team -- the cleanest case of external adoption in this batch. CLUE provides its own traditional-to-simplified conversion tool for models evaluated in simplified Chinese but does not otherwise alter the data.\n",
  "metric": {
   "baseline_note": "93.30 F1 is the human-performance figure from DRCD's own paper (against a baseline-model F1 of 89.59%). This is not a CLUE-run figure: CLUE's live composite leaderboard does not score DRCD at all, so no CLUE-specific human baseline exists for it.\n",
   "direction": "higher_is_better",
   "human_baseline": 93.3,
   "max_score": 100,
   "name": "exact match (EM); F1 also reported by the original paper",
   "unit": "%"
  },
  "name": "CLUE: DRCD (Traditional Chinese span-extraction reading comprehension)",
  "page_kind": "subset",
  "paper": {
   "arxiv": "1806.00920",
   "title": "DRCD: a Chinese Machine Reading Comprehension Dataset",
   "url": "https://arxiv.org/abs/1806.00920",
   "year": 2018
  },
  "publisher": {
   "authors": [
    "Chih Chieh Shao",
    "Trois Liu",
    "Yuting Lai",
    "Yiying Tseng",
    "Sam Tsai"
   ],
   "org": "Delta Research Center -- the dataset's Chinese name (\u53f0\u9054\u95b1\u8b80\u7406\u89e3\u8cc7\u6599\u96c6) identifies it as Delta Electronics'; no further institutional detail was confirmed from the sources read",
   "url": "https://github.com/DRCKnowledgeTeam/DRCD"
  },
  "released": "2018-06",
  "repo_url": "https://github.com/DRCKnowledgeTeam/DRCD",
  "sources": [
   {
    "accessed": "2026-09-08",
    "title": "DRCD: a Chinese Machine Reading Comprehension Dataset (Shao, Liu, Lai, Tseng, Tsai, arXiv:1806.00920)",
    "url": "https://arxiv.org/abs/1806.00920"
   },
   {
    "accessed": "2026-09-08",
    "title": "DRCKnowledgeTeam/DRCD GitHub repository (README)",
    "url": "https://github.com/DRCKnowledgeTeam/DRCD"
   },
   {
    "accessed": "2026-09-08",
    "title": "CLUEbenchmark/CLUE GitHub repository (README, task 8: DRCD)",
    "url": "https://github.com/CLUEbenchmark/CLUE"
   },
   {
    "accessed": "2026-09-08",
    "title": "clue/clue dataset metadata, Hugging Face API (config drcd)",
    "url": "https://huggingface.co/api/datasets/clue/clue"
   },
   {
    "accessed": "2026-09-08",
    "title": "OpenCompass CLUE_DRCD_gen_1bd3c8.py config",
    "url": "https://github.com/open-compass/opencompass/blob/main/opencompass/configs/datasets/CLUE_DRCD/CLUE_DRCD_gen_1bd3c8.py"
   },
   {
    "accessed": "2026-09-08",
    "title": "CLUE1.1 leaderboard, cluebenchmarks.com -- DRCD is not among its scored columns (fetched via scripts/benchmarks/fetch.py)",
    "url": "https://www.cluebenchmarks.com/rank.html"
   }
  ],
  "status": "active",
  "subcategory": "span-extraction reading comprehension (Traditional Chinese)",
  "summary": "CLUE's Traditional-Chinese span-extraction task, adopted unchanged from the separately published Delta Reading Comprehension Dataset; not scored on CLUE's live leaderboard.",
  "tags": [
   "chinese",
   "reading-comprehension",
   "span-extraction",
   "traditional-chinese",
   "clue-subset"
  ],
  "task_format": "Extractive question answering: given a Traditional Chinese passage and question, output the answer text span; the reference OpenCompass config scores it by exact match (EM), and the original paper additionally reports F1.\n"
 }
}