File size: 3,297 Bytes
b36bf78 a1f5a09 b36bf78 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 | ---
pretty_name: FinancialAuditBench
license: mit
language:
- en
tags:
- benchmark
- financial-audit
- agents
- evaluation
configs:
- config_name: completion
default: true
data_files:
- split: test
path: data/completion.parquet
- config_name: review
data_files:
- split: test
path: data/review.parquet
---
# FinancialAuditBench
[Code](https://github.com/modus-audit/financial-audit-bench) · [Trajectories](https://huggingface.co/datasets/modusaudit/FinancialAuditBench-Trajectories)
FinancialAuditBench evaluates AI agents on financial statement audit tasks. It contains
**90 tasks across 6 synthetic audit engagements** in manufacturing and staffing services.
Agents use supporting documents to complete or review spreadsheet workpapers.
## Task settings
| Configuration | Tasks | Starting workpaper | Goal |
| --- | ---: | --- | --- |
| `completion` | 45 | Blank template | Perform the specified procedures and document findings. |
| `review` | 45 | Completed model workpaper | Review the workpaper and correct errors. |
Each configuration has a `test` split with one row per task. Review tasks correspond to
completion tasks and use the same supporting binder. Select either configuration in the
viewer’s subset dropdown.
## Getting started
Install `datasets` and `huggingface_hub`. While the dataset is private, sign in with
`hf auth login` using an account with access.
```python
from datasets import load_dataset
from huggingface_hub import snapshot_download
# Download the task tables, workbooks, and supporting documents.
data_dir = snapshot_download(
"modusaudit/FinancialAuditBench",
repo_type="dataset",
revision="tasks-2026-09-24",
local_dir="benchmark_tasks",
allow_patterns=["README.md", "data/**", "tasks/**", "review_tasks/**", "inputs/**"],
)
# Choose "completion" or "review".
tasks = load_dataset(data_dir, "completion", split="test")
print(tasks[0]["instruction"])
```
## Dataset structure
- `tasks/`: completion instructions, blank workbooks, task configurations, and rubrics.
- `review_tasks/`: review instructions, completed starting workbooks, task configurations, rubrics, and blank grading templates.
- `inputs/`: supporting documents shared by tasks from the same engagement.
- `data/`: completion and review task tables in Parquet format.
Each row includes `task_id`, `base_task_id`, `binder_id`, `industry`, `workpaper`, and
`instruction`, together with links and relative paths to its starting workbook, binder,
and rubric. `base_task_id` connects each review task to its completion task.
See the [benchmark code](https://github.com/modus-audit/financial-audit-bench#run)
for setup, execution, and grading instructions.
## Citation
To reference this dataset release:
```bibtex
@misc{huang2026financialauditbenchbenchmarkconstructiondifferential,
title={FinancialAuditBench: Benchmark Construction under Differential Privacy Using Real-World Priors},
author={Jerry Huang and Sarvesh Babu and Matt Van Buren and Alexander Wang and Pranav Pillai and Arush Jain and James P. Burton and Julia Hockenmaier},
year={2026},
eprint={2609.32835},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2609.32835},
}
```
## License
[MIT License](LICENSE).
|