SpreadsheetBench: SkillOpt mode=multi — 30-turn code repair with execution feedback, 600 s task budget, training-only gold cell feedback f24ba1b verified emrekuruu commited on Sep 6
SpreadsheetBench: actor prompts aligned verbatim with the SkillOpt codegen harness a15f83f verified emrekuruu commited on Sep 5
Argument descriptions carry format only; environment description is the single home of world rules adaffb2 verified emrekuruu commited on Sep 4