README.md

---
license: cc-by-4.0
pretty_name: IndicBankBench
task_categories:
  - text-generation
tags:
  - benchmark
  - banking
  - tool-calling
  - function-calling
  - agents
  - india
  - arxiv:2609.29167
language:
  - en
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: preview/cases.jsonl
---

# IndicBankBench — A Benchmark for Evaluating the Safety and Reliability of Language Models in Indian Retail Banking

[📄 Paper](https://arxiv.org/abs/2609.29167) · [Code](https://github.com/npci/IndicBankBench)

IndicBankBench evaluates whether language models behave safely and reliably in Indian
retail-banking interactions. Its 799 synthetic, multi-turn cases test grounding in customer
context, safe action-taking with mocked banking tools, appropriate clarification and refusal,
and complete resolution of customer requests. All data is synthetic and contains no real customer
information.

## Layout

Each case is a JSON file at `case_bank/<domain>/<tool>/<axis>.NNN.json`. Key fields: `case_id`,
`axis`, `domain`, `tool`, `target_behavior`, `setup` (persona + `login_context` +
`prior_messages`), `tools_exposed`, `user_turns`, `mock`, `gold`, and `grading`.

The Dataset Viewer reads the generated `preview/cases.jsonl`; the harness uses the canonical JSON
case files.

## Running with the harness

Clone the [IndicBankBench](https://github.com/npci/IndicBankBench) harness, then download the
cases and run:

    # Clone and install
    git clone https://github.com/npci/IndicBankBench.git
    cd IndicBankBench
    python3 -m venv .venv
    source .venv/bin/activate
    python -m pip install --upgrade pip
    python -m pip install -e .

    # Copy the configuration template
    cp .env.example .env

    # Download the case data
    hf download NPCI/IndicBankBench --repo-type dataset --local-dir ./data
    export INDICBANKBENCH_DATA=./data

    # Run an evaluation
    python -m harness.cli run --run-id my_model_v1

Set the candidate and judge endpoints in `.env`. See `meta.json` in this repo for the checksums
of this snapshot, including the content-only checksum recorded by the harness.

## Scoring model

Every case is graded in four phases — **S** (safe), **A** (actions), **R** (response), **Q**
(quality) — a case passes only if every active S, A, and R gate holds. S and A are deterministic
code checks; R is the single LLM-judged gate.

## Citation

```bibtex
@misc{paul2026indicbankbench,
  title         = {IndicBankBench: Evaluating Safety and Reliability of Language Model Assistants in Indian Retail Banking},
  author        = {Suvradip Paul and Chandra Bhushan and Harsh Sharma and Nitin Kukreja and Yatharth Dedhia and Keyur Doshi and Prashant Devadiga},
  year          = {2026},
  eprint        = {2609.29167},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  doi           = {10.48550/arXiv.2609.29167},
  url           = {https://arxiv.org/abs/2609.29167}
}
```

## License

The case data is licensed under CC BY 4.0 (see `DATA_LICENSE.md`). See
`DISCLAIMER.md` for the full disclaimer.