Source inventory
Counts below are retained decisions, not original corpus sizes or independent documents. Train refers to CJ-Dataset; test refers to CJ-Bench. Validation is additional and is counted in each repository's manifest. The profiles record terms found during construction; a repository's code license does not establish a separate data license. Source-specific attribution and use conditions continue to apply.
General
Twenty public sources and a programmatic supplement. ASAP and Dianping sentiment share the dianping_reviews family in the benchmark manifest, but are separated here by task.
| Source | Train | Test | Recorded license / terms |
|---|---|---|---|
| T2Ranking | 2,328,371 | 7,809 | Apache-2.0 |
| ASAP | 1,318,889 | 14,290 | Apache-2.0 |
| Dianping sentiment | 1,256,330 | 4,480 | research-only |
| JD full reviews | 973,902 | 3,329 | research-only |
| DMSC | 933,902 | 3,329 | CC0-1.0 |
| MEP-3M | 898,261 | 3,151 | academic-research-only |
| ChID | 635,043 | 3,152 | Apache-2.0 |
| CMNLI | 587,172 | 12,112 | academic-research-only-CNSD-original-terms |
| Ifeng news | 504,284 | 3,151 | research-only |
| CPED (text) | 150,465 | 6,304 | Apache-2.0 |
| DuReader Yes/No | 130,717 | 7,632 | LUGE-noncommercial-research-and-teaching |
| CMRC2019 | 93,605 | 3,048 | CC-BY-SA-4.0 |
| OCNLI | 59,158 | 0 | CC-BY-NC-2.0 |
| PAWS-X (Chinese) | 45,508 | 1,970 | Google-dataset-free-use-with-attribution-request |
| Classical Poetry Retrieval | 29,796 | 13,316 | CC-BY-NC-SA-4.0 |
| LogiQA 2 (Chinese) | 11,526 | 1,573 | CC-BY-NC-SA-4.0 |
| C3 | 11,502 | 3,152 | non-commercial-research |
| Explicit general rules | 10,700 | 859 | Project-generated; separate redistribution license not specified |
| MASSIVE (Chinese) | 10,675 | 2,966 | CC-BY-4.0 |
| ChnSentiCorp | 7,187 | 1,177 | LUGE-noncommercial-research-terms |
| USTS (TED-X) | 3,007 | 3,200 | MIT |
Medical
The 30 retained source families are listed below using their provenance IDs. Some families are task subsets rather than whole upstream datasets. chip_mdcfnpc and cmedcausal use the AlternateCD research subsets; psyqa uses public examples. All Score decisions come from cblue_qtr. The 44-candidate CTC/CMeIE, 148-candidate TCM-SD, and 120-candidate LCMDC fine-label tasks did not pass the retained input budget; LCMDC's 14-class task remains.
| Source family | Train | Test | Recorded license / terms |
|---|---|---|---|
| cblue_qic | 6,211 | 1,693 | Verify original Tianchi download/data agreement; repository code Apache-2.0 does not establish dataset permissions |
| cblue_qqr | 13,004 | 1,492 | Verify original Tianchi download/data agreement; repository code Apache-2.0 does not establish dataset permissions |
| cblue_qtr | 13,281 | 2,913 | Verify original Tianchi download/data agreement; repository code Apache-2.0 does not establish dataset permissions |
| cblue_sts | 3,558 | 1,988 | Verify original Tianchi download/data agreement; repository code Apache-2.0 does not establish dataset permissions |
| chip_cdn | 216,891 | 29,687 | source dataset terms; code license does not establish separate data license |
| chip_mdcfnpc | 27,636 | 1,926 | PromptCBLUE academic research only; reformatted author research release |
| cmb | 213,190 | 6,491 | Apache-2.0 release; retain exam provenance |
| cmedcausal | 3,356 | 1,315 | PromptCBLUE academic research only; reformatted author research release |
| cmedqa2 | 1,486,404 | 58,100 | Non-commercial research only; original dataset README |
| cmeee | 55,121 | 9,811 | source dataset terms; code license does not establish separate data license |
| cmexam | 57,588 | 4,445 | Apache-2.0 code; README academic/research-only data use |
| cmid | 8,906 | 1,113 | Research use per upstream; verify complete data-use statement before use |
| cnmleqa | 3,710 | 464 | CC-BY-4.0 |
| dialmed | 17,348 | 1,072 | MIT repository; preserve original medical data provenance |
| dxy | 332 | 100 | Original research release; independent data license not specified |
| empec | 71,747 | 2,062 | Public author release; independent data license not specified |
| explain_cpe | 5,817 | 187 | Data publicly released for research; no independent LICENSE confirmed |
| gmd12 | 1,732 | 238 | Original research release; independent data license not specified |
| imcs21 | 129,261 | 19,306 | Original research release; independent data license not specified |
| lcmdc | 313,515 | 27,628 | CC-BY-4.0 |
| meddg | 165,337 | 6,634 | GPL-3.0 repository; independent data license not specified |
| medical_ds | 470 | 130 | Original research release; independent data license not specified |
| medqa_mainland | 19,489 | 2,438 | MIT repository; preserve original medical examination provenance |
| medqa_taiwan | 9,852 | 1,316 | MIT repository; preserve original medical examination provenance |
| mlec_qa | 69,902 | 6,806 | MIT repository; preserve exam provenance |
| psy_insight | 11,667 | 974 | Original research release; independent data license not specified |
| psyqa | 709 | 89 | public research examples; complete data requires signed author agreement |
| remedi | 22,023 | 5,161 | MIT |
| text2dt | 1,842 | 585 | source dataset terms; code license does not establish separate data license |
| webmedqa | 50,101 | 3,836 | Repository Apache-2.0; retain original question-answer provenance and data-use terms |
Legal
| Source | Train | Test | Recorded license / terms | Supervision |
|---|---|---|---|---|
| AppealCase | 8,462 | 500 | CC BY-NC 4.0 | Claim-support states; mixed upstream annotations |
| ClaimGen-CN | 8,400 | 200 | CC BY-NC-SA 4.0 | Original cause-of-action classification |
| LeCaRDv2 | 3,000 | 739 | MIT release | Graded case relevance |
| Explicit legal rules | 2,200 | 1,498 | Project-original, Apache-2.0 | Fictional rules with programmatically checked answers |
| LCR-CN | 1,200 | 909 | CC BY 4.0 | Norm conflicts; clause-level and reserved-document panels |
| eCAIL | 400 | 0 | MIT repository release | Ordered matching derived from legal-element overlap |
| CAPP-130 | 338 | 54 | CC BY 4.0 data declaration | Policy topics, importance, risk, and closed-set counts |
| ELAM | 0 | 400 | MIT dataset release | Held-out case matching |
Finance
| Source | Train | Test | Recorded license / terms | Supervision |
|---|---|---|---|---|
| CFLUE | 8,000 | 1,200 | CC BY-NC-SA 4.0 | Examination answer keys |
| FinTruthQA | 6,000 | 900 | Apache-2.0 | Question status, answer relevance and readability |
| Explicit financial rules | 5,000 | 750 | Project-original, CC0-1.0 | Fictional rules with programmatically checked answers |
| FinRE | 3,000 | 450 | MIT | Directed entity-relation labels |
| RoleCS (finance) | 2,000 | 300 | MIT release | Strategy labels in upstream model-generated dialogues |
Interpretation
The labels measure the source tasks: retrieval relevance, answer-key membership, claim support, and readability are distinct. They are not interchangeable measures of factual correctness or real-world safety. Source transformations and parent identifiers are recorded in the released cases where available. Cross-domain and model-pretraining overlap are not comprehensively established.