pointerbench-sheets/REPRODUCE.md

Reproducing Pointerbench-Sheets

The benchmark is generated programmatically from the spreadsheet-coordinate-grounding generator in the source repository.

# from the generator directory in the source repo:
python3 build_pointerbench.py --n 500 --seed 90111 --out /path/to/pointerbench-sheets

• Seed 90111 is disjoint from the training-data seed (7), so every benchmark sheet is unseen during training.
• The build renders one distinct sheet per example, saves it as data/test/<i>.png, and writes one data/test/metadata.jsonl row with the instruction, the pixel bbox, the center point, and the category / UI-style / language tags.
• The benchmark includes ordinary object-like targets (cells, headers, colored regions) plus thin spreadsheet edge targets: column resize handles, row resize handles, right cell edges, bottom cell edges, and the four corners of visible cells. It also includes relative-navigation tasks such as "three rows below A5" and "two columns right and one row down from B7".
• Rendering is fully deterministic given the seed (pure PIL), so the box coordinates are pixel-exact and the set is byte-reproducible.

Regenerating with the same seed reproduces the identical 500 examples. To grow or refresh the set, change --n / --seed and bump the dataset version.