xhcloud's picture
Roll back to 2026-09-29 state (commit 7d47e72): restore v2 splits/indexes/sidecar, remove v3-only files (v3 moved to search_jev_unified_v3_release_20261002)
cbbb487 verified
|
Raw History Blame Contribute Delete
21.3 kB

search_jev_unified_v2 数据报告(2026-09-29 更新)

对象:本目录 train/val/test.jsonl + *_index.jsonl(正式版,897,035 行)。 全部数字由全量扫描得出(扫描前已按 manifest.json 复核六文件 SHA-256), 机器可读版本见 data_report.json。

1. 总量

切分 行数 题数 choice score noul
train 487,012 683,315 204,609 255,190 223,516
val 159,124 200,228 48,329 52,639 99,260
test 250,899 295,675 83,850 36,253 175,572
合计 897,035 1,179,218 336,788 344,082 498,348
  • train 平均 1.40 题/行(qd_quality 每行 2–5 题、helpsteer2 5 维,其余多为 1–2 题)。
  • 2026-09-29 屏蔽了不可判定题(见第 10 节第 5 条),题数因此比 09-28 版少 38,963 道,行数少 2,013 行。
  • 输入级检查:三切分内同 (state, questions) 重复 0、同输入标签冲突 0。
  • 指令多样化行:train 252,752 / val 31,996 / test 41,328(索引带 instruction_variant)。

2. 任务族

task_family train val test 合计 train core train aux
retrieval_judgment 153,990 43,446 18,080 215,516 66,482 87,508
query_understanding 74,318 48,013 45,211 167,542 64,527 9,791
evidence_judgment 62,482 20,328 17,708 100,518 45,568 16,914
general_understanding 61,511 6,413 6,437 74,361 – 61,511
search_action 56,981 10,640 15,793 83,414 51,886 5,095
query_assessment 44,592 25,194 23,543 93,329 – 44,592
query_classification 33,138 5,090 11,461 49,689 1,764 31,374
sealed_evaluation – – 112,666 112,666 – –
  • train:core 230,227(47.3%)/ aux 256,785(52.7%)。val/test 行无 training_lane。
  • sealed_evaluation(只在 test):qrecc_test_equivalence 28,222、candy_evidence 39,989、search_arena_intent 20,127、 ragtruth 17,721、frames 5,976、agentrewardbench 631。它们是封存官方评测源,训练中无同源数据,单独报告。
  • val 中 query_understanding 占 30%(atec 19,357 行),test 中 sealed_evaluation 占 45%; 不要直接看 val/test 总体准确率,按族 / 来源分层汇总。

按 input_origin:

input_origin train val test 合计
search_pool(搜索主池) 212,839 111,537 198,786 523,162
search_hardneg 86,024 8,506 8,153 102,683
query_assessment 44,592 25,194 23,543 93,329
general_mix 61,511 6,413 6,437 74,361
jev_extra(qd_quality_v1) 50,672 2,478 2,615 55,765
query_addon(query 分类补充) 31,374 4,996 11,365 47,735

3. 逐来源

难度列 = 均值(档位范围);置信度列 = 该来源全部题 confidence 均值。语言为索引 language_heuristic 主值。

task_family source capability train lane label_tier 语言 train val test 难度 置信度
evidence_judgment evidence/cfever_ev evidence aux derived_evidence zh 6,935 0 0 3.00 (3) 0.850
evidence_judgment evidence/trendfact_verdict evidence core source_verdict zh 5,847 318 334 3.18 (3–4) 0.900
evidence_judgment evidence/vitaminc evidence core source_verdict en 25,069 0 0 2.00 (2) 0.900
evidence_judgment evidence/vitaminc_conflict evidence aux derived_conflict en 9,979 18,628 16,338 3.00 (3–4) 0.850
evidence_judgment search_hardneg/hn_suff evidence_sufficiency core derived_llm_filtered en 14,652 1,382 1,036 4.05 (4–5) 0.905
general_understanding classification/amazon_reviews_multi_zh review_rating aux original_label zh 4,000 400 400 3.00 (3) 0.970
general_understanding classification/chnsenticorp sentiment_binary aux original_label zh 3,963 400 400 1.00 (1–2) 0.970
general_understanding classification/clue_cmnli nli aux original_label zh 3,000 400 400 3.00 (3) 0.900
general_understanding classification/clue_csl keyword_match aux original_label zh 3,000 400 400 1.96 (1–2) 0.970
general_understanding classification/clue_ocnli nli aux original_label zh 5,000 400 400 3.00 (3) 0.970
general_understanding classification/clue_tnews topic_classification aux original_label zh 5,000 400 400 2.00 (2) 0.970
general_understanding classification/clue_wsc coreference aux original_label zh 752 60 101 3.00 (3) 0.970
general_understanding classification/multi_emotion_dialogue emotion aux original_label zh 2,892 115 111 2.00 (2) 0.900
general_understanding classification/online_shopping_10cats sentiment_binary aux original_label zh 3,959 397 400 1.00 (1–2) 0.970
general_understanding classification/thucnews_title topic_classification aux original_label zh 3,000 400 400 2.00 (2) 0.970
general_understanding classification/weibo_senti_100k sentiment_binary aux original_label zh 5,000 400 400 1.00 (1) 0.970
general_understanding qa/clue_c3 reading_mc aux original_label zh 3,000 400 400 3.43 (3–4) 0.970
general_understanding qa/logiqa_zh logic_mc aux original_label zh 2,000 400 400 4.00 (4) 0.970
general_understanding scoring/coig_p preference_pair aux original_label zh 3,000 400 400 3.64 (3–4) 0.970
general_understanding scoring/cvalues_zh preference_pair aux original_label zh 3,000 400 400 3.00 (3–4) 0.970
general_understanding scoring/douban_reviews_1to5 review_rating aux original_label zh 4,000 400 400 3.00 (3) 0.970
general_understanding scoring/helpsteer2_0to4 response_quality aux original_label en 2,000 229 208 4.77 (4–5) 0.970
general_understanding scoring/helpsteer3_chinese preference_pair aux original_label zh 945 12 17 3.74 (3–5) 0.970
general_understanding scoring/jd_review_1to5 review_rating aux original_label zh 4,000 400 400 3.00 (3) 0.920
query_assessment query_assessment/query_info query_info aux llm_vote / ladder_construct / structural_anchor zh 14,764 9,819 9,139 2.66 (2–3) 0.869
query_assessment query_assessment/query_need query_need aux llm_vote / structural_anchor / llm_generated zh 17,914 7,602 6,974 3.00 (3) 0.882
query_assessment query_assessment/query_noise query_noise aux real_query_rule / synthetic_rule / llm_generated zh 11,914 7,773 7,430 2.00 (2) 0.953
query_classification classification/banking77 query_classification aux original_intent_transfer en 9,973 0 3,071 2.00 (2) 0.970
query_classification classification/clinc150 query_classification aux original_intent_transfer en 10,594 3,092 5,499 2.00 (2) 0.970
query_classification classification/massive_zh query_classification aux original_intent_transfer zh 10,807 1,904 2,795 2.00 (2) 0.970
query_classification decision/kuake_qic intent core original_intent zh 1,764 94 96 2.00 (2) 0.970
query_understanding retrieval/afqmc intent_equivalence core original_binary zh 10,847 3,070 0 2.00 (2) 0.970
query_understanding retrieval/atec intent_equivalence core original_binary zh 8,849 19,357 20,000 2.00 (2) 0.970
query_understanding retrieval/bq_corpus intent_equivalence core original_binary zh 10,792 8,703 8,819 2.00 (2) 0.970
query_understanding retrieval/chip_sts intent_equivalence core original_binary zh 7,771 0 0 2.00 (2) 0.970
query_understanding retrieval/dialogue_rewrite_equivalence rewrite aux weak_negative en 5,177 1,836 1,960 1.50 (1–2) 0.700
query_understanding retrieval/lcqmc intent_equivalence core original_binary zh 18,287 8,389 12,465 2.00 (2) 0.970
query_understanding retrieval/paws_x_zh intent_equivalence core original_binary zh 7,981 1,830 1,967 2.00 (2) 0.970
query_understanding retrieval/qrecc_train_equivalence rewrite aux weak_negative en 4,614 4,828 0 1.86 (1–3) 0.700
retrieval_judgment retrieval/duretrieval relevance aux derived_qrels zh 5,918 780 642 1.88 (1–3) 0.900
retrieval_judgment retrieval/kuake_qqr relevance core native_grade zh 6,866 0 0 2.00 (2) 0.950
retrieval_judgment retrieval/kuake_qtr relevance core native_grade zh 7,882 0 0 3.00 (3) 0.950
retrieval_judgment retrieval/multi_cpr_ecom relevance aux derived_qrels zh 7,493 2,970 0 1.67 (1–2) 0.900
retrieval_judgment retrieval/multi_cpr_medical relevance aux derived_qrels zh 7,375 3,998 0 1.69 (1–2) 0.900
retrieval_judgment retrieval/multi_cpr_video relevance aux derived_qrels zh 7,399 2,683 0 1.66 (1–3) 0.900
retrieval_judgment retrieval/qbqtc relevance core native_grade zh 17,594 18,734 3,681 2.00 (2) 0.950
retrieval_judgment retrieval/qd_quality_v1 qd_quality aux llm_teacher zh 50,672 2,478 2,615 4.35 (4–5) 0.831
retrieval_judgment retrieval/t2reranking relevance core original_binary zh 14,654 3,582 3,370 2.14 (2–3) 0.914
retrieval_judgment retrieval/t2retrieval relevance aux derived_qrels zh 8,651 6,255 5,868 2.00 (1–3) 0.900
retrieval_judgment search_hardneg/hn_gain retrieval_gain core derived_llm_filtered en 19,486 1,966 1,904 3.18 (3–4) 0.874
sealed_evaluation decision/agentrewardbench sealed_evaluation – official_eval en 0 0 631 3.97 (3–4) 0.970
sealed_evaluation decision/search_arena_intent sealed_evaluation – official_eval en 0 0 20,127 2.00 (2) 0.970
sealed_evaluation evidence/candy_evidence sealed_evaluation – official_eval zh 0 0 39,989 2.00 (2) 0.970
sealed_evaluation evidence/ragtruth sealed_evaluation – official_eval en 0 0 17,721 3.92 (3–4) 0.970
sealed_evaluation retrieval/frames sealed_evaluation – official_eval en 0 0 5,976 3.00 (3) 0.970
sealed_evaluation retrieval/qrecc_test_equivalence sealed_evaluation – official_eval en 0 0 28,222 1.86 (1–3) 0.854
search_action decision/xyz_aquila_action action_imitation aux weak_demonstration en 5,095 5,482 10,580 2.98 (2–4) 0.700
search_action search_hardneg/hn_next_hop next_query core derived_llm_filtered en 29,422 2,924 2,923 2.03 (2–3) 0.949
search_action search_hardneg/hn_query_fit query_fit core derived_llm_filtered en 22,464 2,234 2,290 3.04 (3–4) 0.925

query_classification 补充行(banking77 / clinc150 / massive_zh)索引无 language_heuristic,上表语言按源数据集标注。

只有 train、无评估切分的来源:cfever_ev、vitaminc、chip_sts、kuake_qqr、kuake_qtr;这些能力只能在同族其他来源上评估。 afqmc、multi_cpr_* 无 test,banking77 无 val。

4. 标签来源与监督强度

label_tier train val test 合计
original_binary 79,181 44,931 46,621 170,733
official_eval 0 0 112,666 112,666
derived_llm_filtered 86,024 8,506 8,153 102,683
original_label 61,511 6,413 6,437 74,361
derived_qrels 36,836 16,686 6,510 60,032
llm_teacher 50,672 2,478 2,615 55,765
native_grade 32,342 18,734 3,681 54,757
llm_vote 26,617 13,168 11,301 51,086
original_intent_transfer 31,374 4,996 11,365 47,735
derived_conflict 9,979 18,628 16,338 44,945
source_verdict 30,916 318 334 31,568
weak_demonstration 5,095 5,482 10,580 21,157
weak_negative 9,791 6,664 1,960 18,415
real_query_rule 5,981 3,875 3,418 13,274
synthetic_rule 5,683 3,696 3,725 13,104
ladder_construct 4,079 2,687 2,592 9,358
derived_evidence 6,935 0 0 6,935
structural_anchor 1,470 1,349 1,958 4,777
original_intent 1,764 94 96 1,954
llm_generated 762 419 549 1,730
  • train 按 supervision:original 205,714 / derived 161,787 / weak 88,137 / 未标(query_addon)31,374;weak=true 92,937 行(19.1%)。
  • train 中 LLM 参与定标的行(llm_teacher + llm_vote + llm_generated + derived_llm_filtered)共 164,075 行,占 33.7%。

5. 难度

difficulty train val test
1 27,579(5.7%) 8,546 17,973
2 248,649(51.1%) 96,690 144,315
3 132,483(27.2%) 48,283 66,248
4 59,577(12.2%) 3,740 20,490
5 18,724(3.8%) 1,865 1,873

train 各族档位占比:

task_family 1 2 3 4 5
general_understanding 21.2% 22.4% 43.7% 10.0% 2.7%
evidence_judgment – 40.1% 34.7% 24.3% 1.0%
retrieval_judgment 6.2% 41.1% 17.6% 24.4% 10.7%
query_understanding 6.7% 90.9% 2.4% – –
query_assessment – 38.2% 61.8% – –
search_action – 50.5% 48.3% 1.3% –
query_classification – 100% – – –
  • 4–5 档几乎全部来自 qd_quality(4–5)、search_hardneg(hn_suff 4–5、hn_gain/hn_query_fit 3–4)与 helpsteer2/logiqa; val 的 4–5 档只有 5,605 行,4–5 档指标主要看 test(其中 sealed 源 ragtruth/agentrewardbench 贡献 4 档 16,934 行)。
  • 59 个来源中 31 个只有单一档位(如 query_classification、intent_equivalence 全为 2 档),档内无区分度,按难度加权时注意。

6. 置信度(软标签)

按题统计 confidence(索引镜像;noul 题为 max(p, 1−p) 对应的校准值):

区间 train val test
[0.95, 1.00] 282,690(41.4%) 100,049 189,847
[0.90, 0.95) 117,044(17.1%) 35,047 23,616
[0.85, 0.90) 196,745(28.8%) 35,541 33,751
[0.75, 0.85) 65,140(9.5%) 11,367 24,788
[0.60, 0.75) 20,795(3.0%) 17,708 23,222
[0.00, 0.60) 901(0.1%) 516 451

各族题均值(全切分):query_classification 0.970、general 0.961、query_understanding 0.940、sealed 0.941、 query_assessment 0.898、evidence 0.875、retrieval 0.867、search_action 0.849。 最低的层:weak_demonstration / weak_negative 恒为 0.70,llm_teacher 0.61–0.85(均值 0.831), llm_vote 0.45–0.95(均值 0.870)。

7. 标签分布(train)

  • 二分类(noul)大多接近均衡;偏斜较明显的:atec 负例 85.9%、query_info.ambiguous 负例 73.3%、duretrieval / t2retrieval 正例 71.5%、afqmc 负例 68.6%、multi_cpr_* 负例约 66%、cfever_ev 负例 64.8%。search_hardneg 四个集为构造性 50/50。
  • 多分类最高类占比:trendfact_verdict refutes 55.6%、vitaminc supports 50.2%、query_noise「有效查询」63.2%、 query_info.missing_main「无明显缺失」49.8%、qd_quality.query_freshness never 49.9%、kuake_qic「治疗方案」35.2%; hn_next_hop 四个选项位置均衡(最高 25.3%)。无任何 (来源, 题目) 的最高类占比 ≥ 0.9。
  • score 题(按软分 Σ档×概率):qd_quality relevance 均值 1.52、satisfaction 1.15(0–3 档,偏低)、timeliness 1.43(屏蔽无时间信息的题后)、 authority 2.16;kuake_qqr 均值 0.63(0–2,偏低档);jd_review 中位数落在最高档(源分布偏 5 星)。

逐 (来源, 题目) 明细见 data_report.json 的 train_label_balance。

8. 指令多样性(train)

66 个 (来源, 题目) 中:54 个 ≥ 10 种指令,5 个 4 种(search_hardneg 四集、query_noise.noise), 3 个 3 种(query_info.info_level、query_need.need_level、query_noise.noise_type), 4 个 2 种(query_info.ambiguous / info_pair / missing_main、query_need.freshness);单一指令 0 个。 query_assessment 与 search_hardneg 是剩余的低多样性来源。

9. 输入长度

state 序列化(json.dumps(ensure_ascii=False))字符数,全切分:

task_family 行数 p50 p95 max >1024 >2048
general_understanding 74,361 65 1,578 22,140 7.5% 3.6%
evidence_judgment 100,518 375 2,785 5,917 18.8% 11.1%
retrieval_judgment 215,516 283 1,410 4,172 17.6% 1.4%
query_understanding 167,542 55 276 2,740 1.2% 0.1%
query_assessment 93,329 13 53 135 0 0
search_action 83,414 250 3,667 4,074 24.9% 24.5%
query_classification 49,689 33 80 435 0 0
sealed_evaluation 112,666 225 3,598 12,580 21.7% 13.4%

不含 questions 文本;按 token 估算时中文约 1 字 ≈ 1 token、英文约 4 字符 ≈ 1 token。

10. 使用时需要注意

  1. 语言:train 中中文主导 63.7%、拉丁文主导 29.7%;search_action / hn_* / evidence 的 vitaminc 系列全部是英文, search_action 族目前没有中文行。
  2. 评估切分构成不均:val 被 atec(19,357)、qbqtc(18,734)、vitaminc_conflict(18,628)主导;test 45% 为 sealed。 汇总指标请按来源宏平均。
  3. 弱标签层:weak_demonstration(xyz_aquila 示范动作)与 weak_negative(改写等价的构造负例)置信度 0.70, qd_quality 为 LLM 教师标签——按 README「使用建议」降权。
  4. 已移除的冲突样本:同输入但标签不一致的 56 行(train 32 / val 18 / test 6)已整组移除,不在本数据中。
  5. 2026-09-29 修订(软标签不变):
    • 屏蔽不可判定题:删除该题,索引 masked_questions 记录原因;一行里的题全被删掉则整行移除。
      • qd_quality timeliness:无发布时间且无查询时间 23,400 道,无发布时间却标 3 档 569 道。
      • qd_quality query_freshness:同一 (query, query_time) 在不同文档下标签不一致,12,981 道。
      • qrecc:上下文为空且改写引入了新实体的正例,2,013 行,整行移除(train 281 / val 250 / test 1,482)。
    • HelpSteer2 complexity / verbosity 量表按 NVIDIA 原始定义重写。
    • 句对等价任务的 state 改为 {query_a, query_b}(paws_x_zh 为 {text_a, text_b})。
    • vitaminc / cfever_ev 的 state 改为 {claim, evidence},instructions 同步改名。
    • xyz_aquila 删去 oxed{} 格式尾巴。
    • 意图迁移源(query_addon)的索引 gold 由 answers argmax 回填。
  6. 2026-09-29 修订:clue_tnews 类别映射修复:原 label id→类名映射自 105 起错位(CLUE 官方无 105/111), 112 旅游与 116 游戏同被映射为"股票",且缺"房产"类,约 64% 的行类别错误。已按原始数据回查 id、改用官方映射: 改标 3,710 行(train 3,204 / val 243 / test 263),全部 5,800 行的选项扩为 15 类(补"房产"), 概率按原置信度 0.97 重算,索引 gold / original_domain_unverified 同步;行数与其他来源不变。
  7. 2026-09-29 修订:helpsteer3_chinese 偏好方向修复:原转换把 overall_preference>0 当作 response1 更好,与 HelpSteer3 约定(及原始 best_response 字段)相反,导致全部偏好标签反向。已逐行按原始 best_response 定位更好的回复,翻转 974 行(train 945 / val 12 / test 17)的 choice / probabilities 与索引 gold;回复位置、置信度与行数不变。
  8. 2026-09-29 修订:ragtruth / frames 字段修复:ragtruth 原把上文截到 800 字(97% 行被截)、回复截到 1,000 字且丢了任务说明,字段还误名为 question/passage;已按原始 RAGTruth 重建为 {query, context, response} 全文(17,721 行逐行回查、标签一致),ragtruth 不受 4,000 字 state 上限约束(截断会让幻觉标签无从判断),因此 sealed_evaluation 的长度统计上升。frames 的 passage 实为维基百科 URL,改为 {question, wiki_url, wiki_title}(标题由 URL 解码,6 条补协议头,3 条 w.wiki 短链无标题)。题目、答案、索引与行数均不变。