README.md

# AI Spreadsheet Benchmark Dataset

The AI Spreadsheet Benchmark captures 53 realistic spreadsheet prompts spanning analysis, enrichment, visualization, and workbook-management workflows. It is designed to evaluate how spreadsheet copilots behave in situ: Do they write formulas? Do charts stay linked to data? Can the output recompute when numbers change?

- **Paper:** ["The AI Spreadsheet Benchmark: Measuring Dynamic Output in Spreadsheet Assistants"](https://huggingface.co/datasets/rowshq/aispreadsheetbenchmark/blob/main/technical_paper.pdf)
- **Dataset:** `rowshq/aispreadsheetbenchmark`
- **Metrics:** Pass@1, Pass@3, Dynamic Output Rate, Latency

## Accessing the dataset

```python
from datasets import load_dataset

benchmark = load_dataset("rowshq/aispreadsheetbenchmark")
questions = benchmark["questions"]        # prompt text, categories, scoring metadata
```

## Task categories

| Category | Tasks | Examples |
| --- | --- | --- |
| Classic Data Analysis | 37 | YoY growth columns, lookups, joins, dashboards, cohort tables |
| Advanced Analysis | 5 | K-means clustering, forecasting, anomaly detection, custom visualizations |
| Creating Models | 2 | Interactive head-to-head calculators, investment simulators |
| Manage Spreadsheet Elements | 3 | Conditional formatting, sorting, chart styling, sheet setup |
| Arithmetic Operations | 6 | High-precision arithmetic sanity checks |

Each prompt record contains:
- Natural-language instructions
- Category and sub-category labels

## Recommended evaluation protocol

1. **Reset** the workbook to the canonical dataset before each run.
2. **Issue** prompts verbatim. If the assistant asks for clarification, respond neutrally while keeping the task scope fixed.
3. **Assess success** using the published acceptance criteria (execution checks whenever possible).
4. **Assess dynamic output** by perturbing underlying data and verifying the response updates automatically (no pasted values or screenshots).
5. **Measure latency** from prompt submission to assistant completion.
6. **Compute metrics:** Pass@1, Pass@3 (up to three attempts per task), Dynamic Output Rate, mean/median latency.

## Baseline results

Initial evaluation across five assistants (Rows AI Analyst, Excel Copilot, Google Sheets + Gemini, Shortcut, Julius):

| Assistant | Pass@1 | Pass@3 | Dynamic (%) | Mean time (s) |
| --- | --- | --- | --- | --- |
| Rows AI Analyst | 89 | 92 | 74 | 220 |
| Microsoft Excel Copilot | 53 | 64 | 8 | 46 |
| Google Sheets + Gemini | 57 | 64 | 6 | 11 |
| Shortcut | 83 | 83 | 13 | 222 |
| Julius | 75 | 83 | 0 | 30 |

Detailed per-category tables and visualizations appear in the accompanying technical paper.

## Citation

```
@misc{rowshq2025benchmark,
  title  = {The AI Spreadsheet Benchmark: Measuring Intelligence in Spreadsheet Assistants},
  author = {Samagaio, Álvaro Mendes and Cruz, Henrique and Pereira, Humberto Ayres and Schulz, Torben},
  year   = {2025},
  url    = {https://huggingface.co/datasets/rowshq/aispreadsheetbenchmark/blob/main/technical_paper.pdf}
}
```

## Questions & contributions

Open a discussion or issue on the Hugging Face dataset page if you:
- Find discrepancies in acceptance criteria or scoring instructions
- Want to share new assistant baselines or evaluation tooling
- Plan to extend the benchmark with additional domains or datasets

We welcome contributions that improve documentation, acceptance criteria, or reproducibility assets. Reach out via the dataset page to coordinate substantial updates.