
SpreadsheetBench 2
What it measures and how
Fix V2’s 321 tasks and common SWE-agent conditions. Standalone office products and other scaffolds do not represent base models. Correct spreadsheet visualization does not establish broad aesthetics.
How this evidence is used
Under observation: stable JSON and row-level dates exist, but base-model coverage under common conditions is insufficient. Integration awaits measurements of current models.
Limitations and attribution
Under observation: stable JSON and row-level dates exist, but base-model coverage under common conditions is insufficient. Integration awaits measurements of current models.
Data license: The official website footer declares MIT. Dataset and code access points are public.
Results are published by SpreadsheetBench / Renmin University / AfterQuery. Raw scores and the e/acc consensus index use different scales and cannot be added directly.