R&D / Banking AI Evaluation Library / Datasets and sources
τ-bench banking_knowledge domain
A simulated customer service domain where an agent uses tools and searches policy documents, with tasks that have defined outcomes.
Open the publisher page: tau2-bench repository (opens in a new tab)
Dodone links to the publisher and does not host or copy this data. Read the licence on the publisher page before you use it.
About this source
- Publisher
- Sierra Research
- Context
- Simulated customer service with a knowledge base, English
- Version
- CHANGELOG 2.0.0 (19 October 2026)
- Last checked
- 2026-11-03
- Labels or answers
- Each task has a defined expected outcome inside the simulated environment.
Facts stated by the publisher
- The banking_knowledge domain was added in version 1.0.0 with 97 tasks and 698 policy and procedure documents.
- The repository states that 2.0 scores are not comparable with 1.x scores.
Licence, access and privacy
- Licence
- MIT License (code only) (opens in a new tab) (Licence not verified)
- Scope
- The MIT licence covers the code. The repository states no licence for the tasks or the knowledge base, so data reuse is not confirmed.
- Access
- Open. Public GitHub repository. Pin the exact release when you cite it.
- Privacy
- Privacy reviewed: low risk. Simulated customers and invented policy documents.
Limitations
- A simulated bank and its invented policies are not proof that an agent follows a real bank process.
- Scores from different releases must not be compared.
Suggested for these evaluations
- Customer policy enquiry and evidence retrieval Direct fit
Why it helps: It has simulated banking service tasks with a policy knowledge base, tools and defined outcomes.
What it does not validate: Real bank policies and systems. Reference clauses for each answer may not be listed.