WTM-BENCH is a benchmark for evaluating LLM agents on realistic,
multi-artifact spreadsheet automation tasks. Each task pairs a starting
Excel workbook with a natural-language request; the agent must drive the
workbook to a target state through a multi-turn tool-calling loop, writing and
executing real code each turn.
Code, runner, grader, and reproduction rollouts:
https://github.com/prose-ms/wtm-bench (see the benchmark.py harness).… See the full description on the dataset page:
https://huggingface.co/datasets/prose-ms/wtm-bench.
