SpreadsheetBench 2
SpreadsheetBench / Renmin University / AfterQuery / Professional work · Complete a full spreadsheet workload covering financial modelling, workbook fixes and data visualisation.
What it measures, and how
The 321 tasks of V2 and the unified SWE-agent conditions are fixed; results from independent office products and other scaffolds are not attributed to the base model. Correctness of table visualisation is not the same as broad aesthetics.
How this evidence is used
Under observation: stable JSON and per-row dates exist, but base-model coverage under the current unified conditions is insufficient; scores will be connected after current-period models are actually tested.
Limits and data attribution
Under observation: stable JSON and per-row dates exist, but base-model coverage under the current unified conditions is insufficient; scores will be connected after current-period models are actually tested.
Data licence: MIT is declared in the footer of the official website; dataset and code entry points are public.
Scores published by SpreadsheetBench / Renmin University / AfterQuery. Raw scores and the News consensus score use different scales and cannot be added directly.