license: cc-by-3.0
task_categories:
- text-generation
language:
- en
tags:
- multimodal
- agent
- workflow
- spreadsheet
- pdf
- image
- code
- finance
- accouning
modalities:
- text
- spreadsheet
- pdf
- image
- code
configs:
- config_name: Finch_Dataset_All
data_files:
- split: test
path:
- finch_workflows_test.jsonl
Finch: Benchmarking Finance & Accounting Workflows around Multimodal Enterprise Spreadsheets
This repository contains the dataset for Finch, an enterprise-level benchmark for evaluating an agent’s ability to act like a skilled finance & accounting expert on real-world workflows.
- Paper: to be added
- Project Page: https://huggingface.co/datasets/FinWorkBench/Finch
- Code: https://github.com/FinWorkBench
Dataset Description
Finch focuses on composite finance & accounting workflows that span:
data entry/import, structuring/formatting, web search, cross-sheet/file retrieval, calculation, financial modeling, validation, translation, visualization, and reporting.
The workflows are derived from real-world enterprise workspaces (Enron and various institutions/companies such as World Bank, Canada goverment, etc), including:
- Large and messy spreadsheets with multimodal artifacts including text, tables, formulas, charts, pivots, images, etc
- Linked PDFs and documents that provide additional business context
We adopt a three-step workflow labeling process:
- Summarizing workflow types supported by real collaborative enterprise email threads.
- Deriving concrete workflow instances from versioned spreadsheets and related files using LLMs.
- Meticulous expert annotation of instructions and reference outputs, involving hundreds of hours of expert work.
This process yields 172 enterprise-grade workflows—primarily multi-task composite — each with carefully written instructions and aligned input/reference files, capturing the intrinsic complexity, messiness, and multimodality of real-world finance & accounting work. In this release, we provide full annotations for the first 72 workflows, with the remaining 100 to be released in a subsequent update.
Experiment results show that even frontier agents solve fewer than 30% of the workflows, revealing a substantial performance gap for real-world enterprise scenarios.
📁 Dataset Structure
The instruction-tuning corpus is released in JSONL format.
Each line corresponds to one workflow-centric example:
{
"id": "<workflow identifier>",
"instruction_en": "<English task instruction for a finance & accounting workflow>",
"source_files": ["<input file name>", "..."],
"source_files_urls": ["<input file download URL>", "..."],
"reference_outputs": {
"files": ["<reference output file name>"],
"text": "<textual reference output>"
},
"reference_file_urls": ["<reference output file download URL>"],
"task_type": "<task category (e.g., reporting, modeling)>",
"business_type": "<business domain (e.g., budgeting, trading)>"
}