ytz009/CwI-Bench

CwI-Bench — Code-with-Image Bench 30 code-with-image tasks where models must write and run code against the image to reach the answer — pure visual inspection ("just look") is insufficient by design: answers require pixel-level precision (±px coordinates, exact counts, sub-degree angles, per-channel color recovery). Each task is a config; each config has three splits: split size source photos role train 60 (15/tier) COCO train2017 reflection / skill-learning (tr60)… See the full description on the dataset page: https://huggingface.co/datasets/ytz009/CwI-Bench.