NPCI/tau-indian-banking

TauIndianBankBench: an Indian retail-banking benchmark for tool-using conversational agents

Data files for the indian_banking domain of the tau2-bench harness: a conversational agent staffs the chat desk of a synthetic India-based retail bank. The code that runs and scores the benchmark lives in the companion repository npci/TauIndianBankBench; this dataset holds only the domain data it loads.

Viewer note. The dataset viewer shows preview/tasks_preview.parquet, which holds the same 1,000 tasks as the canonical tau2/domains/indian_banking/tasks.json that the benchmark loads.

Contents

tau2/domains/indian_banking/tasks.json          1,000 tasks across 50 scenario families
tau2/domains/indian_banking/split_tasks.json    task-id splits (base = all 1,000)
tau2/domains/indian_banking/db.json             synthetic bank database (944 customers)
tau2/domains/indian_banking/policy.md           the bank's written policy shown to the agent
tau2/domains/indian_banking/agent_instruction.txt   agent system instruction
tau2/domains/indian_banking/kb/                 24 knowledge-base documents + provenance.json
tau2/user_simulator/simulation_guidelines.md    user-simulator guideline

Each task carries the customer persona and scenario, the required tool actions in order (with argument checks), the information the agent must communicate, natural-language assertions, and the checks that decide the reward. Episodes are scored on the final database state and on the required tool actions; the communication and assertion checks are recorded alongside.

Usage

pip install huggingface_hub
python - <<'PY'
from huggingface_hub import snapshot_download
snapshot_download("NPCI/tau-indian-banking", repo_type="dataset", local_dir="data")
PY

Place the files under the repository's data/ directory (the companion repository's scripts/fetch_data.py does exactly this), then run the benchmark as described in its README.

Synthetic data notice

Every customer, account, card, deposit, loan, transaction, address, e-mail address and phone number is generated. Bank, brand, merchant, employer and locality names are invented; any resemblance to real persons, institutions or places is coincidental. Knowledge-base documents paraphrase publicly available regulatory and product information for the sole purpose of grounding the simulated agent and are not advice.

Disclaimer

This dataset is provided for research, testing and benchmarking only, on an "as is" basis and without warranties of any kind. Nothing in it is legal, regulatory, compliance, financial, banking, risk, security or professional advice, and it should not be relied upon for operational, customer-facing or production decisions. The full terms, including the limitation of liability and the position on trademarks, are in DISCLAIMER.md.

Licence

MIT. Authors: Krishanu Adhikary, Prashant Devadiga (NPCI AI Research Team).

Citation

If you use this data, please cite the benchmark together with the tau2-bench and tau-bench papers it builds on (CITATION.cff in the companion repository carries the same entries).

@software{npci2026tauindianbankbench,
  title   = {TauIndianBankBench: an Indian retail-banking benchmark for tool-using conversational agents},
  author  = {Adhikary, Krishanu and Devadiga, Prashant},
  organization = {NPCI AI Research Team},
  year    = {2026},
  version = {1.0.0},
  note    = {A $\tau^2$-bench domain: 1000 tasks across 50 scenario families, 34 tools, strict end-state scoring}
}

@misc{barres2025tau2bench,
  title         = {$\tau^2$-Bench: Evaluating Conversational Agents in a Dual-Control Environment},
  author        = {Victor Barres and Honghua Dong and Soham Ray and Xujie Si and Karthik Narasimhan},
  year          = {2025},
  eprint        = {2506.07982},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2506.07982}
}

@misc{yao2024taubench,
  title         = {$\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains},
  author        = {Shunyu Yao and Noah Shinn and Pedram Razavi and Karthik Narasimhan},
  year          = {2024},
  eprint        = {2406.12045},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2406.12045}
}