BedBench

sample_id string	query_image_1 image	query_image_2 image	answer bool
0			true
1			false
2			true
3			true
4			false
5			false
6			false
7			true
8			false
9			true
10			true
11			false
12			false
13			false
14			true
15			false
16			true
17			false
18			true
19			false
20			true
21			true
22			true
23			false
24			false
25			true
26			false
27			false
28			false
29			true
30			true
31			true
32			false
33			false
34			true
35			false
36			true
37			true
38			true
39			false
40			false
41			false
42			false
43			false
44			true
45			false
46			true
47			false
48			true
49			true
50			false
51			false
52			false
53			true
54			true
55			true
56			true
57			false
58			true
59			false
60			true
61			false
62			true
63			false
64			true
65			true
66			true
67			true
68			false
69			false
70			false
71			true
72			true
73			false
74			false
75			true
76			true
77			true
78			true
79			false
80			false
81			true
82			true
83			false
84			true
85			false
86			false
87			false
88			true
89			false
90			true
91			true
92			false
93			false
94			true
95			false
96			false
97			true
98			false
99			false

BedBench

Evaluation data, model outputs, and probing features for the paper Where Do Vision-Language Models Fail at 3D Reasoning? Every task is posed on rendered views of 100 procedurally generated voxel objects (T1-T12) or 100 multi-object scenes (T13). Code (generator, evaluation harness, probing, fine-tuning) is released separately.

Configurations

Each configuration is one task under one visual condition (the default setting is listed first).

Task	Configurations
T1	`uncolor`, `uncolor_axes`, `color`
T2	`uncolor`, `uncolor_axes`
T3	`uncolor_axes`, `color_axes`
T4	`color`, `color_axes`
T5	`color`, `color_axes`
T6	`uncolor`, `uncolor_axes`
T7	`color`, `color_axes`
T8	`uncolor`, `uncolor_axes`
T9	`uncolor`, `uncolor_axes`
T10	`color`, `color_axes`
T11	`invalid_uncolor`, `invalid_uncolor_axes`, `invalid_color`, `valid_uncolor`, `valid_uncolor_axes`, `valid_color`
T12	`uncolor`, `uncolor_axes`, `color`
T13	`invalid_uncolor`, `invalid_color`, `valid_uncolor`, `valid_color`

T5_finetune holds the 800 training and 100 validation objects of the T5 fine-tuning experiment (generator seed 100; no object overlaps the benchmark).

from datasets import load_dataset
ds = load_dataset("wnkh/BedBench", "T5_color", split="train")

The evaluation harness reads this repository directly: python -m bed_eval run --dataset wnkh/BedBench ....

Extras

File	Contents
`extras/samples.tar.gz`	the 100 objects (manifest with all targets, every rendered view, projections, distractors) and the 100 T13 scenes, as written by the generator
`extras/outputs.tar.gz`	every model response used in the paper (12 open-weight models x 2 modes x 5 runs, 4 frontier models, image-removal, answer-format, reasoning-budget, factorized-T5, and fine-tuned runs) with ground truth;`bed_eval score outputs` reproduces the paper's numbers
`extras/probing_features.tar`	mean-pooled features of the 14 frozen encoders used in the probing analysis

Developer

Main developer: Nguyen Kim Hai Bui (GitHub).

License

CC BY 4.0.

Downloads last month: 102