SpreadsheetBench
Real-world spreadsheet manipulation benchmark with 912 authentic questions for LLMs.
Open dataset Updated 2026-06-16
What is SpreadsheetBench?
SpreadsheetBench is a benchmark of 912 real questions for spreadsheet manipulation, collected exclusively from practical user workflows and paired with corresponding spreadsheet files.
It supports evaluation of large language models on authentic spreadsheet tasks and is intended for researchers comparing model performance against existing synthesized benchmarks.
What you can build with SpreadsheetBench
LLM spreadsheet agent evaluation
Test how well a new LLM or agent handles real user queries like formula creation, data filtering, and chart generation on authentic .xlsx files.
Fine-tuning for office automation
Use the 912 question-file pairs as supervised data to fine-tune models that output correct spreadsheet operations or Python code for pandas/openpyxl.
Comparative benchmarking
Run standardized evaluations against other LLMs to measure progress on practical spreadsheet manipulation without relying on synthetic test sets.
Load SpreadsheetBench
Python
from datasets import load_dataset
ds = load_dataset("KAKA22/SpreadsheetBench")- 1pip install datasets
- 2from datasets import load_dataset
- 3dataset = load_dataset('kaka22/SpreadsheetBench')
- 4Access 'questions' and linked spreadsheet files in the dataset splits
- 5Run your model on each query and compare outputs against ground-truth actions
SpreadsheetBench: pros & cons
Pros
- +Uses unaltered real-world spreadsheet files
- +912 questions from actual user scenarios
- +Avoids synthetic data and oversimplified structures
- +Directly measures practical LLM manipulation skills
Cons
- –Only 912 examples total
- –Requires handling complex real spreadsheet formats
- –Primarily designed for LLM evaluation rather than general ML training
Did you find this helpful?
Frequently asked questions
A benchmark with 912 real-world spreadsheet questions that tests LLMs on manipulation tasks using actual unaltered files.
User reviews
Verified reviews from the community shape this listing's rating.
Loading reviews…
Similar datasets
Other ai & machine learning options worth comparing.