What’s Out There
← Back to swiping
Internet & techFreeAcademic

SpreadsheetBench

Renmin University of China (RUC)

912 real Excel-forum questions, each tested on multiple spreadsheet files, to measure LLM spreadsheet skills.

A benchmark built from questions asked on online Excel forums, with their spreadsheets. One row is an instruction with its spreadsheet files and expected outputs. Evaluation resembles an online judge: each instruction has about three test files, so a solution must work generally.

Access

Cost
FreeNo charge.
License
Open licenseCC BY-SA 4.0
Sign-up
No sign-up
How you get it
Bulk downloadJSON, XLSX, Tar archive

What's in it

One row is
One row per instruction with test spreadsheets
Covers
Not geographic
Years
Forum posts (see paper)
Updated
One-time release
Size
912 instructions and 2,729 test cases
instructioninstruction typespreadsheet fileanswer positiontest cases

Quality

4 of 5

Real questions with robust multi-file tests, and a successor version exists.

  • Forum spreadsheets may contain personal or business data
  • Download is a tar archive rather than a Parquet table

Connects to

  • GDPval: Broader professional-task benchmark that includes spreadsheet deliverables
  • GAIA: Assistant benchmark that sometimes uses spreadsheet attachments

Questions it could answer

  • Which instruction types have the lowest success rates?
  • How much does a model's result change on different test spreadsheets?

Similar data sets

  • MMLU-Pro

    A harder MMLU with ten answer options and 12,000 reasoning-heavy questions.

    EducationTIGER-Lab (University of Waterloo and collaborators)Free

  • Humanity's Last Exam

    Thousands of expert-written questions designed to be too hard for current AI, across many fields.

    EducationCenter for AI Safety and Scale AIFree

  • OpenR1-Math-220k

    220,000 math problems, each with two to four DeepSeek-R1 reasoning traces, many checked automatically.

    EducationHugging Face (Open-R1 project)Free

  • MMMLU (Multilingual MMLU)

    The MMLU test set translated into 14 languages by professional human translators.

    EducationOpenAIFree

  • SWE-bench Verified

    500 real GitHub issues, human-checked, used to score whether AI can fix bugs in popular Python projects.

    Internet & techPrinceton NLP (with OpenAI)Free

  • Llama-Nemotron Post-Training Dataset

    Millions of math, code, and reasoning examples used to post-train NVIDIA's Llama-Nemotron models.

    Language & textNVIDIAFree