anonymity11/FormulaCascade_Benchmark

FormulaCascade Benchmark

FormulaCascade is a benchmark for long-horizon spreadsheet formula-dependency reasoning in real executable .xlsx workbooks. It evaluates whether agents can recover hidden workbook logic by localizing relevant regions, tracing formula dependencies, reconstructing latent formulas or values, and filling masked target cells.

Current Review Release

This repository currently hosts the standard eval-200 split used for reproducible model comparison. The split is a stratified subset of the full 11,171-task FormulaCascade benchmark:

• 200 tasks total: 50 tasks per tier.
• Balanced sheet topology: 25 single-sheet and 25 cross-sheet tasks per tier.
• Four quality-difficulty tiers: T1_FRONTIER, T2_HARD, T3_MEDIUM, and T4_STANDARD.
• Executable artifacts: solver-visible masked workbooks plus evaluator-only contracts for hidden targets, formulas, cached values, dependency layers, and provenance.

Uploaded files include:

• FormulaCascade_Benchmark_eval_200.tar.gz: compressed eval-200 split.
• FormulaCascade_Benchmark_eval_200/: expanded eval-200 directory for direct browsing.
• data/eval_200_manifest.parquet: lightweight tabular manifest for the Hugging Face Dataset Viewer.
• data/eval_200_manifest.csv and data/eval_200_manifest.jsonl: equivalent manifest exports for convenience.
• SHA256SUMS.txt: checksum for release integrity verification.

The full benchmark contains 11,171 bilingual tasks across finance, sales, operations, engineering, HR, education, and general analytics workbooks, with dependency depth up to 46 and substantial cross-sheet structure. The full release uses the same task schema, hidden-contract format, and tiered directory layout as eval-200.

Dataset Viewer

The Dataset Viewer displays data/eval_200_manifest.parquet, not the raw .xlsx files. Spreadsheet workbooks and hidden contracts are stored as repository files and should be downloaded or browsed directly.

Manifest columns include:

• fcb_id: stable task identifier.
• tier / tier_label: quality-difficulty tier.
• category: single_sheet or cross_sheet.
• domain: workbook domain label.
• depth: dependency-chain depth.
• is_cross_sheet: whether the task requires cross-sheet reasoning.
• total_masked: number of masked cells.
• formula_families: coarse formula-family labels.
• masked_workbook: path to the solver-visible workbook.
• task_description: path to the natural-language task description.
• evaluator_contract_dir: path to evaluator-only ground-truth contracts.

Usage

Download the archive:

tar -xzf FormulaCascade_Benchmark_eval_200.tar.gz

For each task, provide the solver with:

• release/FCB-NNNNN/task_description.json
• release/FCB-NNNNN/task-NNNNN.xlsx

Use the corresponding internal/FCB-NNNNN/ directory for evaluation only. It contains target contracts, dependency layers, ground truth, provenance, and release-quality audit files.

Intended Use

FormulaCascade is intended for evaluating spreadsheet agents, LLM tool-use systems, formula-reasoning methods, and spreadsheet-native harnesses under controlled executable conditions. It is designed for research, benchmarking, diagnostics, and reproducibility studies.

Out-of-Scope Use

FormulaCascade is not intended for training production models on evaluator-only hidden contracts, making decisions about individuals, or evaluating unrestricted spreadsheet UI automation, macro execution, or non-spreadsheet office workflows.

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).