---
license: other
pretty_name: FxBench
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- spreadsheets
- excel
- formula-completion
---

# FxBench

FxBench is a spreadsheet formula-completion benchmark. Each row contains
one target formula and the corresponding workbook snapshot needed to solve it.

## Dataset contents

This release contains 503 examples.

Columns in the Hugging Face Parquet split:

| Column | Type | Description |
| --- | --- | --- |
| `id` | string | Public example ID in the form `fxbench-{i}`. |
| `function` | string | Primary Excel function for the target formula. |
| `formula` | string | Ground-truth Excel formula. Empty for `ABSTAIN` rows where no formula response is expected. |
| `target_cell` | string | Cell address where the formula belongs. |
| `target_sheet` | string | Worksheet containing the target cell. |
| `workbook` | binary | XLSX workbook snapshot bytes. |

The auxiliary `metadata.jsonl` manifest contains scalar public metadata only; workbook bytes are in the Parquet split.
The loadable split is `data/train-00000-of-00001.parquet` and stores the `workbook` column as XLSX bytes.

Workbook bytes are sanitized before export to remove document properties,
classification-label metadata, custom properties, comment/person metadata, and
local path markers while preserving workbook sheets and links.

## Loading

```python
from datasets import load_dataset

ds = load_dataset("<your-namespace>/<your-dataset-repo>", split="train")
row = ds[0]

with open(f"{row['id']}.xlsx", "wb") as f:
    f.write(row["workbook"])
```