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)
validation	50 (13/13/12/12)	COCO train2017	selection set (sel50, the exact stratified slice used in runs)
test	100 (25/tier)	COCO val2017	final evaluation (test_mini)

Zero photo overlap between train/validation (train2017) and test (val2017). Difficulty tiers: easy / medium / hard / expert in the difficulty column.

from datasets import load_dataset
d = load_dataset("ytz009/image_pot_bench", "alignment")
item = d["test"][0]   # columns: id, image, image_2, prompt, answer, difficulty, meta

image_2 is non-null only for two-image tasks (e.g. congruent_pair). meta is a JSON string with generation parameters (constructive ground truth). Scoring rules per task are given in the Scoring section below.

Scoring

Every item is scored all-or-nothing (0/1) — there is no partial credit, and a task's accuracy is the mean over its 100 held-out test items. The model's answer is the last <answer>…</answer> span in its output; no span scores 0. The rule column below is the exact condition under which a prediction is counted correct (it scores 1 if the condition holds, else 0). Discrete answers (permutations, counts, index sets, transform tuples) must match exactly; continuous answers (pixels, angles, RGB values, percentages) must fall within the task's tolerance — the tolerance is part of the task definition, and every generator rejects any sample a shortcut could land inside it.

Across the 30 tasks: 9 exact · 20 tolerance · 1 exact+tolerance (ring_rotor).

Each entry reads task — criterion — the condition under which a prediction is counted correct.

Transformation

• circular_shift — tolerance — dx,dy each within ±2 px of the true shift
• restore_then_count — exact — the integer equals the true count
• shift_rot_align — tolerance — dx,dy within ±2 px each and θ within ±1.875° (circular)
• alignment — tolerance — dx,dy each within ±2 px
• rot_scale — tolerance — angle within ±3° (circular) and scale within ±0.05
• transform_chain — exact — the (dx,dy,rot_k,fliph) tuple matches in every field
• unshuffle2d — exact — both the row and column permutations match exactly
• jigsaw — exact — the full piece-order sequence matches exactly
• rotated_jigsaw — exact — the full tile-and-rotation assignment matches exactly
• restore_then_measure — tolerance — the percentage is within ±1.5 pp
• ring_rotor — exact+tol — the ring index is exact and the angle within ±5°
• double_exposure — tolerance — every signed shift is within ±8 px (count and order match)

Perception

• grating_period — tolerance — the period is within ±1.5
• anomaly_v2 — exact — the reported cell set equals the true anomaly set
• peak_count — exact — the integer equals the true count
• congruent_pair — exact — the reported index pair equals the true congruent pair
• trace_length — tolerance — length/200 is within ±0.3
• scan_moire — tolerance — the scale is within ±0.004

Localization

• crop_provenance — tolerance — the box reaches IoU ≥ 0.5 with the true crop (coords normalized 0–1000)
• brightest_blob — tolerance — the point is within ±2 px on the 480-px canvas
• sub_rect_swap — tolerance — both boxes reach IoU ≥ 0.5, order-free (coords normalized 0–1000)
• retouch_region — tolerance — the box reaches IoU ≥ 0.5 with the true region (coords normalized 0–1000)
• circle_fit — tolerance — the point is within ±4 px on the 480-px canvas
• print_misreg — tolerance — dx,dy are each within ±0.3 px
• gain_center — tolerance — every spotlight centre is within ±15 px (in strength order)

Color

• film_color — tolerance — R,G,B each within ±10
• swatch_rgb — tolerance — R,G,B each within ±10
• patch_mean_dots — tolerance — R,G,B each within ±10
• color_count — exact — the integer equals the true count
• channel_mix — tolerance — R,G,B each within ±8

Tasks (30)

Transformation & rearrangement: circular_shift, restore_then_count, shift_rot_align, alignment, rot_scale, transform_chain, unshuffle2d, jigsaw, rotated_jigsaw, restore_then_measure, ring_rotor, double_exposure

Perception & counting: grating_period, anomaly_v2, peak_count, congruent_pair, trace_length, scan_moire

Localization: crop_provenance, brightest_blob, sub_rect_swap, retouch_region, circle_fit, print_misreg, gain_center

Color: film_color, swatch_rgb, patch_mean_dots, color_count, channel_mix

Note: trace_length uses the revised (v2fix) generator with a reference-solver fairness screen.