| --- |
| license: cc-by-4.0 |
| pretty_name: IndicBankBench |
| task_categories: |
| - text-generation |
| tags: |
| - benchmark |
| - banking |
| - tool-calling |
| - function-calling |
| - agents |
| - india |
| language: |
| - en |
| size_categories: |
| - n<1K |
| --- |
| |
| # IndicBankBench dataset |
|
|
| IndicBankBench is a 799-case benchmark for evaluating tool-using language-model assistants in |
| Indian retail banking. Each JSON case describes a synthetic, scripted, multi-turn interaction, |
| the available mock tools, the expected tool behavior, and staged grading criteria. It contains |
| no real customer records. |
|
|
| Anonymous-review locations: |
|
|
| - Dataset: https://huggingface.co/datasets/anonuser73268/IndicBankBench |
| - Dataset ID: `anonuser73268/IndicBankBench` |
| - Compatible code: https://github.com/anonuser73268/IndicBankBench |
|
|
| ## Files and split |
|
|
| The release has one unsplit case bank: |
|
|
| ```text |
| case_bank/<domain>/<tool>/<axis>.NNN.json |
| ``` |
|
|
| There are 799 JSON files and 799 unique `case_id` values. Cases are grouped by domain and |
| primary tool for navigation; those directories are not training, validation, or test splits. |
| File order is the lexicographic order of relative paths, which is also the order used by the |
| reference harness. |
|
|
| - `case_bank/`: canonical case JSON files |
| - `SCHEMA.md`: field and format documentation |
| - `meta.json`: count, layout, compatibility, and aggregate checksum metadata |
| - `VALIDATION.md`: release validation results and known linter findings |
| - `LICENSE`: CC BY 4.0 license notice and attribution guidance |
| - `DISCLAIMER.md`: intended-use and warranty disclaimer |
|
|
| ## Load the cases |
|
|
| The files can be loaded with the Python standard library: |
|
|
| ```python |
| import json |
| from pathlib import Path |
| |
| root = Path("case_bank") |
| cases = [json.loads(path.read_text(encoding="utf-8")) |
| for path in sorted(root.rglob("*.json"))] |
| assert len(cases) == 799 |
| assert len({case["case_id"] for case in cases}) == 799 |
| ``` |
|
|
| To use the reference harness from a sibling checkout: |
|
|
| ```bash |
| cd ../code |
| export INDICBANKBENCH_DATA=../data |
| python -m unittest discover -s indicbankbench/tests |
| python indicbankbench/scripts/case_lint.py |
| ``` |
|
|
| ## Dataset characteristics |
|
|
| - Language: English |
| - Modality: scripted multi-turn text interactions plus structured mock-tool data |
| - Domains: five operational banking domains plus capability/refusal cases |
| - Tools: 32 mock banking tools |
| - Primary evaluation axes: 20 |
| - Personal data: synthetic fixtures only |
|
|
| The dataset is intended for controlled research, testing, and benchmarking. It must not be used |
| for live banking, customer, legal, regulatory, credit, fraud, or financial decisions. |
|
|
| ## Provenance |
|
|
| The cases and bank-specific fixture fields are synthetic. Names and selected demographic |
| descriptors used when constructing the synthetic login contexts were derived from |
| [NVIDIA Nemotron-Personas-India](https://huggingface.co/datasets/nvidia/Nemotron-Personas-India), |
| which is licensed under CC BY 4.0. Bank-specific identifiers, masked contact fields, addresses, |
| dates, and relationship fields were generated synthetically. |
|
|
| ## License and citation |
|
|
| The case data and authored dataset documentation are licensed under Creative Commons Attribution |
| 4.0 International; see [LICENSE](LICENSE). Evaluation results and model or judge outputs are not |
| part of this dataset release. During anonymous review, cite the accompanying anonymous paper |
| rather than attempting to identify its authors. Final citation metadata will be added after |
| review. |
|
|
