README.md

---
license: cc-by-3.0
tags:
- agent
- workflow
- multimodal
- spreadsheet
- pdf
- image
- code
- finance
- accounting
- tabular
modalities:
- text
- spreadsheet
- pdf
- image
- code
- tabular

configs:
  - config_name: Finch_Dataset_All
    data_files:
    - split: test
      path:
        - finch_workflows_test.jsonl
---

<img src="figs/finch_workflow.jpeg" width="1000" />

# Finch (FinWorkBench): Benchmarking Finance & Accounting across Spreadsheet-Centric Enterprise Workflows

This repository contains the dataset for **Finch**, an enterprise-grade benchmark for evaluating an agent’s ability to work like a skilled finance & accounting expert (work IQ) on real-world professional workflows.

* **Paper**: https://arxiv.org/abs/2512.13168
* **Evaluation Code**: https://github.com/FinWorkBench/Finch

---

## 🍻 Updates

* **2026-4-6**: FinWorkBench is accepted to ACL 2026 Findings.

---

## Dataset Description

Finch focuses on **messy and long-horizon 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—primarily Enron and the EUSES Corpus—and further extended with more recent artifacts (e.g., 2024 and 2025) from investment and securities companies, the World Bank, Canadian and British government agencies, and other sources.
- Enterprise **email threads** where collaborators naturally describe, discuss, and track workflows
- Large and messy **spreadsheets** with multimodal artifacts including text, tables, formulas, charts, pivots, images, etc 
- Interlinked **PDFs and documents** that provide additional business context  

We adopt a three-step workflow labeling process:

1. **Inducing workflow types and instances** from real collaborative context in **enterprise email threads** ([Enron Corpus](https://en.wikipedia.org/wiki/Enron_Corpus): 500,000 emails from 150 executives and employees).  
2. **Deriving concrete workflow instances** by analyzing changes across **spreadsheet versions** (15,000 versioned spreadsheet files from Enron and [EUSES](https://dl.acm.org/doi/10.1145/1082983.1083242)) and designing workflows based on high-quality recent artifacts from investment and securities companies, World Bank, Canadian/British government agencies, WideSearch, Dabstep, and more.  
3. **Conducting meticulous expert annotation** of task instructions, input files, and reference outputs, involving hundreds of hours of expert work.

<img src="figs/annotation.png" width="1000" />

This process yields **172 enterprise-grade workflows—primarily multi-task composite workflows**, involving 1,710 spreadsheets and 27 million cells, capturing the intrinsic **messy, long-horizon, knowledge-intensive, and collaborative nature** of real-world finance & accounting work. 92.4% of workflows involve multiple input and reference sheets, with an average of 8 sheets and a long tail reaching 91 spreadsheets. In this release, we provide full annotations for the 172 workflows and the evaluation code.

<img src="figs/distribution_chart.png" width="1000" />

We conduct both human and automated evaluations of frontier AI systems including GPT5.1, Claude Sonnet 4.5, Gemini 3 Pro, Grok 4, and Qwen 3 Max. GPT 5.1 Pro spends 16 mins for each workflow on average and gets the highest pass rate of 38%, while Claude Sonnet 4.5 passes just 25.0%, revealing a substantial performance gap for real-world enterprise scenarios.

<img src="figs/exp_results.jpg" width="1000" />

---

## 🔍 Why FINCH is Hard

Most of the individual capabilities Finch probes — reading tables, interpreting formulas, writing code, searching the web — are things frontier LLMs already appear to handle well on isolated benchmarks. Yet performance degrades sharply on Finch. Our analysis points to **five intertwined properties** of real-world enterprise F&A work that make failures more likely and more catastrophic:

1. **Large, fragmented spreadsheet ecosystems.** Workflows routinely span dozens of interlinked workbooks and thousands of rows distributed across many sheets. Executing them accurately requires long-range cross-sheet navigation and precise referencing, which substantially increases the likelihood of small retrieval errors.

2. **Dense, semantically homogeneous content.** Many cells contain domain-specific financial concepts that are subtly different yet lexically similar (e.g., variants of revenue/expense items, adjusted vs. unadjusted metrics), making entity disambiguation and cell grounding unusually difficult.

3. **Complex and often irregular layouts.** Multi-level headers, merged cells, nested subtotals, and bespoke layouts force models to infer structure from noisy contents and ad hoc formatting. Tiny misinterpretations (e.g., off-by-one errors when specifying ranges) can propagate into globally incorrect outputs, especially when the same logic is applied in batch across many sheets.

4. **Latent business logic encoded in formulas.** Formulas encode temporal assumptions, fine-grained dependencies, and business logic that is not visible from displayed values alone. For example, a column header `IF NGPL MidContinent index (@ Baker)` looks like a daily exposure metric, but the associated formula `25 * V21 + C41 * C22` actually encodes a 55-day payment timing. Models that prioritize cell values and under-use formulas systematically misinterpret such columns, and the error propagates through subsequent steps.

5. **Multimodal, cross-artifact reasoning.** Many workflows combine spreadsheets with PDFs, charts, Word documents, and screenshots, requiring the agent to jointly reason over heterogeneous formats. Tables embedded in PDFs, for instance, are often only partially referenced, with key entries missing from the text channel.

It is the **combination** of these factors — rather than any single one — that drives the sharp performance degradation on real enterprise workflows, and it translates into substantial multi-step agent interaction at execution time (quantified below).

---

## 📊 Operational Complexity

Even successful Finch workflows demand substantial multi-step agent interaction. We conduct a case study using **Claude Coworker (Opus 4.6)** on 20 representative tasks (full table in the paper appendix). Excluding two web-search-heavy workflows that require 71 and 107 tool calls respectively — dominated by `websearch` / `webfetch` for external evidence gathering — the remaining tasks range from **6 to 25 tool calls** (mean 13.2, median 14, IQR 9–17).

Two observations stand out:

- **Web-grounded tasks incur dramatically higher overhead.** On the two outlier tasks, web search and fetch together account for 48–56% of all tool calls; external retrieval, cross-source comparison, and verification dominate cost.
- **Task count alone does not predict overhead.** Even two workflows with the same task-type label (e.g., both `Calculation`) can differ nearly 2× in tool calls — intrinsic business-logic reasoning (sign conventions, method selection, cross-sheet consistency) matters more than how many task types a workflow is tagged with.

In other words: even single- or two-task workflows can require a dozen or more interleaved read / compute / verify steps, so the practical complexity of Finch is substantially higher than task counts alone would suggest.

---

## Examples

Example 1: Review the Inv & WC Value Adj summary tab and add the missing cross‑sheet data references to the other worksheets so the roll‑up pulls the correct figures. Return the updated file with those links in place.

<img src="figs/example_1.jpeg" width="1000" />

Example 2: Add a new worksheet named "Scenario3" to the same workbook, mirroring the structure, row/column layout, monthly detail table, and chart area of "Scenario1". For Scenario3, update the hedging assumptions to a balanced allocation: 10-Yr 25%, 5-Yr 20%, 1-Yr 15%, May-Sep 20%, Q3 15%. Keep the note "Maximum Monthly Average Short Position to Cover (July Peak) = 30,508 MW" unchanged; only the new sheet should be added, and formulas may be used within it.

<img src="figs/example_2.jpeg" width="1000" />

Example 3: Transcribe the content from the image into the Excel file.

<img src="figs/example_3.jpeg" width="1000" />

Example 4: Per the red parameters and the Method 1/Method 2 guidance noted in H8 and H9, complete the formulas in columns T and U (starting from1), Note that the starting point for the formulas in columns T and U is 1, representing the initial signal to hold Index 1. In the formulas for columns T and U, 1 represents the signal to hold Index 1, -1 represents the signal to hold Index 2, and 0 represents the signal to make no change. Then complete column I. The method selection in B11 should drive the model so that all cells and charts refresh consistently when switching between methods.

<img src="figs/example_4.jpeg" width="1000" />

---


## 📁 Dataset Structure

The corpus is released in **JSONL** format.  
Each line corresponds to one **workflow-centric example**:

```json
{
  "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)>",
  "task_constraints": "<task constraints (e.g., perform modifications rather than generation from scratch)>"
}
```

Annotations of Finch should **NOT** appear in training corpora.

---

## 📣 Feedback & Issues

If you find any issues with the dataset or have suggestions, please open a discussion in the **Community** tab — we value your feedback!

## Citation
```bibtex
@article{dong2025finch,
  title={Finch: Benchmarking Finance \& Accounting across Spreadsheet-Centric Enterprise Workflows},
  author={Dong, Haoyu and Zhang, Pengkun and Gao, Yan and Dong, Xuanyu and Cheng, Yilin and Lu, Mingzhe and Yakefu, Adina and Zheng, Shuxin},
  journal={arXiv preprint arXiv:2512.13168},
  year={2025}
}
```

**📧 Contact:** donghaoyu82@gmail.com