RLVR data
Bank-operations tasks with verifiable rewards. Each task carries a machine-checkable outcome, so a policy can be scored without a human grader.
- Deterministic checkers per task
- Synthetic ledgers, accounts and counterparties
Coming soon!
FinExhaust builds training and evaluation material for agents that work inside bank operations, including the exceptions and edge cases. Everything is generated against a synthetic institution, so no customer data is involved.
Bank-operations tasks with verifiable rewards. Each task carries a machine-checkable outcome, so a policy can be scored without a human grader.
Multi-step agent traces that include the failure and the repair: wrong tool call, bad assumption, caught exception, corrected path.
BankingEnv (cleanroom-eval) is an MIT-licensed evaluation set with 192 public development cases and 192 held-out cases. Cases are synthetic and machine-checked.
Commercial licences for the RLVR and SFT material. A sample pack is available under NDA so your team can inspect format and quality before any commitment.
Questions about data, evaluation or licensing go to Paul Farrington.
Ficta Meridian (fictameridian.com) is FinExhaust's synthetic demo institution. It is not a real bank and holds no real accounts.
This page sets no cookies and loads no analytics or third-party scripts.