Datasets:
Tasks:
Question Answering
Formats:
parquet
Languages:
English
Size:
< 1K
Tags:
finance
structured-finance
asset-backed-securities
sec-edgar
long-context
document-understanding
License:
Add eval.yaml: register WaterfallBench as a Hub benchmark (Harbor)
Browse files
eval.yaml
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: WaterfallBench
|
| 2 |
+
description: >
|
| 3 |
+
WaterfallBench tests whether a model can execute an asset-backed securities waterfall to the cent from the deal
|
| 4 |
+
documents alone. Each of the 20 tasks is a real auto-loan or lease securitization at its first distribution date:
|
| 5 |
+
the final prospectus, the closing agreements and that month's servicer report with every waterfall result removed
|
| 6 |
+
(7-11 SEC EDGAR filings, 500K-740K tokens). The model returns the servicing fee, each note class's interest,
|
| 7 |
+
principal and ending balance, the reserve balance and the residual in integer cents, graded exactly against what
|
| 8 |
+
the trustee paid. Tasks run as Harbor tasks; answer keys are private. Scores are percentages (0-100).
|
| 9 |
+
evaluation_framework: harbor
|
| 10 |
+
|
| 11 |
+
leaderboard:
|
| 12 |
+
base_models_only: false
|
| 13 |
+
|
| 14 |
+
tasks:
|
| 15 |
+
# Primary metric: share of tasks where every graded field is exact to the cent.
|
| 16 |
+
- id: fully_resolved
|
| 17 |
+
config: default
|
| 18 |
+
split: test
|
| 19 |
+
# Share of non-trivial graded fields exactly right.
|
| 20 |
+
- id: field_accuracy
|
| 21 |
+
config: default
|
| 22 |
+
split: test
|