Concerns about the SpreadsheetBench Verified

[huntzhan]

@Kaka23333 thank you for releasing SpreadsheetBench Verified, I have some questions/concerns:

1. I noticed that in previous versions each sample had three test cases, but now there's only one. Is this by design? Related to this, I saw shortcut.ai mentioned avg4=86pp and bo4=98.2pp in their blog post (https://www.tryshortcut.ai/blog/posts/spreadsheetbench-verified). These metrics seem surprisingly high, and I'm wondering if the reduction from three test cases to one has significantly decreased the difficulty of the modified benchmark. This could potentially reduce the benchmark's influence in the long run, as it may become less attractive for researchers to work with if it's not challenging enough. Are there any SOTA model metrics available for comparison?

2. I see a new data_position field in dataset.json. Is this field only used during the annotation and data filtering process, or will it also be used in inference prompts? (especially by shortcut.ai?)