fxBench: Evaluating and Understanding Formula Suggestions in Spreadsheets
Paper • 2610.24729 • Published
Datasets:
id stringlengths 9 11 | function stringlengths 1 16 | formula stringlengths 0 559 | target_cell stringlengths 2 5 | target_sheet stringlengths 1 31 | workbook unknown |
|---|---|---|---|---|---|
fxbench-1 | ABS | =ABS(ATAN((B7-B6)/(C7-C6))*180/PI()) | G7 | Dis, Bear, N&E | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
46,
204,
204,
41,
99,
1,
0,
0,
86,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
209,
78,
... |
fxbench-2 | ABS | =ABS(B2-AVERAGEIFS(B$2:B$11,$A$2:$A$11,$A2)) | G2 | Sheet1 | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
46,
204,
204,
41,
99,
1,
0,
0,
86,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
209,
78,
... |
fxbench-3 | ABS | =ABS(AVERAGE(C9:C10)-C9)*B9 | D9 | Sheet1 | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
46,
204,
204,
41,
99,
1,
0,
0,
86,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
209,
78,
... |
fxbench-4 | ABS | =ABS(X8*COS(DEGREES(U7-X7))) | X9 | Performance | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
217,
156,
228,
28,
103,
1,
0,
0,
227,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
81,
79,... |
fxbench-5 | ABS | =ABS(IFERROR((J3/$D3),0)) | O3 | 1 | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
46,
204,
204,
41,
99,
1,
0,
0,
86,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
209,
78,
... |
fxbench-6 | ABS | =ABS(IF(E2<0,E2,"")) | G2 | Month 01 | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
46,
204,
204,
41,
99,
1,
0,
0,
86,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
209,
78,
... |
fxbench-7 | ABS | =ABS(MAX(Q6:S6)-MIN(Q6:S6)) | V6 | TABLE | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
217,
156,
228,
28,
103,
1,
0,
0,
227,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
81,
79,... |
fxbench-8 | ABS | =ABS(MAX(F2:F7)-MIN(H2:H7)) | V2 | Sheet1 | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
46,
204,
204,
41,
99,
1,
0,
0,
86,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
209,
78,
... |
fxbench-9 | ABS | =ABS(MAX(D8:D17)-MIN(D8:D17))*1000 | D18 | X-Y-Z REP | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
160,
180,
1,
138,
114,
1,
0,
0,
138,
8,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
150,
77,
110,... |
fxbench-10 | ABS | =ABS(MAX(I3:I35)-MIN(I3:I35))*12 | I38 | Trials | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
46,
204,
204,
41,
99,
1,
0,
0,
86,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
209,
78,
... |
fxbench-11 | ABS | =ABS(B32-MIN(IF(D9:D20>B32,D9:D20))) | B37 | Sheet1 | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
46,
204,
204,
41,
99,
1,
0,
0,
86,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
209,
78,
... |
fxbench-12 | ABS | =ABS(MOD(A2,1)-MOD(B2,1)) | C2 | Sheet1 | [
80,
75,
3,
4,
20,
0,
0,
0,
8,
0,
0,
0,
33,
0,
46,
204,
204,
41,
99,
1,
0,
0,
86,
6,
0,
0,
19,
0,
0,
0,
91,
67,
111,
110,
116,
101,
110,
116,
95,
84,
121,
112,
101,
115,
93,
46,
120,
109,
108,
205,
149,
209,
78,
... |
End of preview. Expand in Data Studio
FxBench
FxBench is a benchmark for evaluating large language models in the inline spreadsheet formula-completion setting. Each row contains one target formula and the corresponding sanitized workbook snapshot needed to solve it.
- Code: https://github.com/SANKET7738/fxbench
- Paper: fxBench: Evaluating and Understanding Formula Suggestions in Spreadsheets (PDF)
- 🎉 NeurIPS 2026: Accepted paper and poster
Dataset contents
This release contains 503 examples in the train 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 stored in data/train-00000-of-00001.parquet.
Loading
from datasets import load_dataset
ds = load_dataset("sanket-mhatre/fxbench", split="train")
row = ds[0]
with open(f"{row['id']}.xlsx", "wb") as f:
f.write(row["workbook"])
Provenance and processing
FxBench is derived from Sheetpedia. FxBench selects and transforms workbook content, defines formula-completion targets, sanitizes workbook metadata, and adds benchmark annotations.
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.
Intended use and limitations
FxBench is intended for research on spreadsheet formula suggestion and evaluation. It contains a fixed sample of spreadsheet contexts and should not be treated as representative of every spreadsheet domain, language, or formula pattern. A model's score does not guarantee that its formulas are correct or safe for production spreadsheets.
License
FxBench is licensed under CC BY-SA 4.0. It is derived from Sheetpedia, which uses the same license. When redistributing or adapting FxBench, provide attribution, link to the license, indicate changes, and share adaptations under the same or a compatible license.
Citation
If you use FxBench, please cite the paper:
@misc{mhatre2026fxbench,
title = {{fxBench}: Evaluating and Understanding Formula Suggestions in Spreadsheets},
author = {Sanket Mhatre and Sumit Gulwani and Vu Le and Yasharth Bajpai and Gust Verbruggen},
year = {2026},
eprint = {2610.24729},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2610.24729}
}
- Downloads last month
- 70