WaterfallBench / eval.yaml
ritik12's picture
Add eval.yaml: register WaterfallBench as a Hub benchmark (Harbor)
4205985 verified
Raw History Blame Contribute Delete
1.04 kB
name: WaterfallBench
description: >
WaterfallBench tests whether a model can execute an asset-backed securities waterfall to the cent from the deal
documents alone. Each of the 20 tasks is a real auto-loan or lease securitization at its first distribution date:
the final prospectus, the closing agreements and that month's servicer report with every waterfall result removed
(7-11 SEC EDGAR filings, 500K-740K tokens). The model returns the servicing fee, each note class's interest,
principal and ending balance, the reserve balance and the residual in integer cents, graded exactly against what
the trustee paid. Tasks run as Harbor tasks; answer keys are private. Scores are percentages (0-100).
evaluation_framework: harbor
leaderboard:
base_models_only: false
tasks:
# Primary metric: share of tasks where every graded field is exact to the cent.
- id: fully_resolved
config: default
split: test
# Share of non-trivial graded fields exactly right.
- id: field_accuracy
config: default
split: test