README.md

---
license: cc-by-nc-4.0
task_categories:
- table-question-answering
language:
- en
- zh
tags:
- spreadsheets
- xlsx
- formula-reasoning
- formula-dependency
- formula-family
- benchmark
- agentic-evaluation
size_categories:
- n<1K
pretty_name: FormulaCascade Benchmark
configs:
- config_name: default
  data_files:
  - split: test
    path: data/eval200_manifest.parquet
---

# FormulaCascade Benchmark

This repository hosts the anonymous review release of FormulaCascade, a benchmark for long-horizon spreadsheet formula-dependency and formula-family reasoning in real `.xlsx` workbooks.

## Current Review Release

The current upload contains the eval-200 release version selected from v5 Challenge/Core/Broad release strata:

- `FormulaCascade_Benchmark_eval_200_Release_version.tar.gz`: compressed eval-200 release-version split.
- `FormulaCascade_Benchmark_eval_200_Release_version/`: expanded eval-200 release-version split for direct browsing in the Hugging Face file tree.
- `data/eval200_manifest.parquet` and `.csv`: viewer-friendly task-level manifest for the Hugging Face Dataset Viewer.
- `FormulaCascade_Benchmark_eval_200_Release_version/eval200_manifest.jsonl` and `.csv`: task-level selection metadata with local paths removed.
- `FormulaCascade_Benchmark_eval_200_Release_version/eval200_metadata.json`: aggregate split metadata.
- `SHA256SUMS.txt`: checksums for integrity verification.
- `CROISSANT_RAI_METADATA.json`: Croissant metadata with minimal Responsible AI fields.

To unpack:

```bash
tar -xzf FormulaCascade_Benchmark_eval_200_Release_version.tar.gz
```

The extracted folder is:

```text
FormulaCascade_Benchmark_eval_200_Release_version/
```

## Split Composition

| Group | Tasks | Single | Cross | Avg depth | Max depth | Avg masked |
|---|---:|---:|---:|---:|---:|---:|
| Challenge | 100 | 50 | 50 | 3.32 | 38 | 1,291.9 |
| Core eval subset | 100 | 50 | 50 | 2.86 | 27 | 546.0 |
| Total | 200 | 100 | 100 | 3.09 | 38 | 918.9 |

The complete v5 strata are Challenge (100 tasks), Core (600 tasks), and Broad (11,304 tasks). This repository provides the official balanced eval-200 release for review and reproducible comparison.

## Directory Structure

- `release/FCB-*/task-*.xlsx`: solver-visible masked workbook.
- `release/FCB-*/task_description.json`: solver-visible task instruction.
- `internal/FCB-*/*`: evaluator-only contracts, ground truth, provenance, and source files. These files must not be visible to solver agents during evaluation.

## Intended Use

FormulaCascade evaluates agents that must localize workbook context, trace formula dependencies, infer formula-family patterns, edit masked workbooks, and produce executable formulas or values. The main deterministic metrics are exact formula pass and executable value pass.