{
 "body": "\nPart of the [FewCLUE](fewclue.md) family.\n\n## What it measures\n\nBUSTM (XiaoBu Dialogue Short Text Matching) gives the model two short, colloquial Chinese sentences\nand asks whether they express the same intent -- a binary match/no-match judgment. The pairs come\nfrom real intent-recognition logs behind XiaoBu (Breeno), OPPO's voice assistant for its phones and\nIoT devices, so the language is conversational rather than written prose (for example, \"\u5973\u5b69\u5b50\u5230\u5e95\u662f\u4e0d\u662f\u4f60\"\npaired with \"\u4f60\u4e0d\u662f\u5973\u5b69\u5b50\u5417\" is judged a match). BUSTM is one of the three FewCLUE tasks built new for the\nbenchmark rather than resampled from the CLUE benchmark.\n\n## Reading the numbers\n\nThe labelled public test set holds 1,772 pairs, scored by accuracy; a majority-class guess scores\n50% since the label is roughly balanced. Human evaluators scored 88.0% on this task; the paper's\nstrongest few-shot method here was EFL (which recasts matching as textual entailment) at 71.8%, the\nbest result EFL achieved on any FewCLUE task and clearly ahead of the 56-61% other methods managed on\nBUSTM specifically. That gap is worth checking a reported score against: a matching score much closer\nto 70% than to the 50-60% range suggests an entailment-style approach rather than plain cloze-style\nprompting. See the [FewCLUE](fewclue.md) family page for the shared few-shot protocol, split\nstructure and contamination notes that apply here too.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "fewclue_bustm",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [],
 "page": {
  "category": "reasoning",
  "dataset": {
   "languages": [
    "zh"
   ],
   "license": "",
   "modalities": [
    "text"
   ],
   "public_test_set": true,
   "size": 1772,
   "size_note": "1,772 labelled public test pairs (test_public.json, used for scoring), plus 2,000 in the private test set (original leaderboard only), 32 train and 32 dev examples per split (5 parallel splits, train_0..train_4 / dev_0..dev_4, plus a merged train_few_all/dev_few_all), and 4,251 unlabelled pairs not used for scoring. Figures from the FewCLUE paper's Table 1, matching the GitHub README.\n",
   "splits": "train_0..train_4 (32 each) + train_few_all; dev_0..dev_4 (32 each) + dev_few_all; test_public (1,772, labelled); test (2,000, private); unlabeled (4,251)",
   "url": "https://github.com/CLUEbenchmark/FewCLUE/tree/main/datasets/bustm"
  },
  "freshness": {
   "researched": "2026-09-08",
   "researched_by": "sonnet-5 agent, batch 5, slice B"
  },
  "harness": {
   "opencompass": "FewCLUE_bustm"
  },
  "id": "fewclue_bustm",
  "leaderboard_url": "https://www.cluebenchmarks.com/fewclue.html",
  "lineage": {
   "family": "fewclue"
  },
  "measures": "Two short, colloquial Chinese sentences drawn from a voice assistant's intent-matching logs; the model judges whether they express the same intent, a binary match/no-match decision, learned few-shot from 32 labelled training pairs.\n",
  "metric": {
   "baseline_note": "50.0% is the paper's majority-class baseline; the public test set's two labels are roughly balanced, so this matches the two-way random-guess rate. See the fewclue family page for the top overall few-shot method scores.\n",
   "direction": "higher_is_better",
   "human_baseline": 88.0,
   "max_score": 100,
   "name": "accuracy",
   "random_baseline": 50.0,
   "unit": "%"
  },
  "name": "FewCLUE: BUSTM (Dialogue Short Text Matching)",
  "page_kind": "subset",
  "paper": {
   "arxiv": "2107.07498",
   "title": "FewCLUE: A Chinese Few-shot Learning Evaluation Benchmark",
   "url": "https://arxiv.org/abs/2107.07498",
   "year": 2021
  },
  "publisher": {
   "authors": [
    "Liang Xu",
    "Xiaojing Lu",
    "Chenyang Yuan",
    "Xuanwei Zhang",
    "Huilin Xu",
    "Hu Yuan",
    "Guoao Wei",
    "Xiang Pan",
    "Xin Tian",
    "Libo Qin",
    "Hu Hai"
   ],
   "org": "CLUE team",
   "url": "https://github.com/CLUEbenchmark/FewCLUE"
  },
  "released": "2021-04",
  "repo_url": "https://github.com/CLUEbenchmark/FewCLUE",
  "sources": [
   {
    "accessed": "2026-09-08",
    "title": "FewCLUE: A Chinese Few-shot Learning Evaluation Benchmark (Xu et al., arXiv:2107.07498)",
    "url": "https://arxiv.org/abs/2107.07498"
   },
   {
    "accessed": "2026-09-08",
    "title": "CLUEbenchmark/FewCLUE GitHub repository (task description, dataset statistics)",
    "url": "https://github.com/CLUEbenchmark/FewCLUE"
   },
   {
    "accessed": "2026-09-08",
    "title": "OpenCompass FewCLUE_bustm dataset config",
    "url": "https://raw.githubusercontent.com/open-compass/opencompass/main/opencompass/configs/datasets/FewCLUE_bustm/FewCLUE_bustm_gen.py"
   }
  ],
  "status": "unknown",
  "subcategory": "dialogue short-text semantic matching (binary intent match), few-shot",
  "summary": "FewCLUE's dialogue short-text matching task: judge whether two short colloquial Chinese sentences share the same intent, learned from 32 labelled training pairs.",
  "tags": [
   "chinese",
   "few-shot",
   "fewclue-subset",
   "semantic-matching"
  ],
  "task_format": "Sentence-pair binary classification (match / no match), graded on the single correct label; evaluated from a 32-example few-shot training split, one of five parallel splits FewCLUE provides for this task.\n"
 }
}