CDSBench dataset
What the corpus contains
Every verified historical case expands the benchmark, improves precedent coverage and exposes new failure modes. The figures below are live counts from the database; nothing is rounded up. Coverage limits are stated at the bottom.
Historical cases
30
5 in preparation
Verified questions
296
of 296 on published cases
Supporting evidence links
454
sections a verified label rests on
Model versions tested
9
with an official run
Reference entities
19
7 corporate · 12 sovereign
Instruments
30
Documents
62
306 citable sections
Years covered
2024–2028
published cases
Credit events
By type
Splits
How cases are used
| Development Mastermind may be improved against these | 18 |
| Validation promotion decisions use these | 6 |
| Holdout never tuned on, excluded from precedent retrieval | 6 |
Ground truth
Verification status
| Public outcome verified official outcome statement names the instrument | 0 |
| Deterministically verified two extractors agree; evidence check passes | 296 |
| Multi source verified agreement across independent sources | 0 |
| Unverified shown, never scored | 0 |
| Disputed excluded from scoring | 0 |
Cases · Restructuring
Clear filtersCoverage limits
- Restructuring events dominate: 28 of 30 published cases. Results on other event types rest on fewer questions.
- No published bankruptcy or repudiation moratorium or obligation acceleration or governmental intervention case yet, so nothing on this site measures models on those event types.
- Published cases span 2024–2028; earlier credit events are not yet covered.
- Deterministic rules simplify the standard Deliverable Obligation Characteristics and are provisional until publicly or expert validated. Rules →
- Real historical cases are added only after their documents are ingested and their labels independently verified; 5 are in preparation.