Datasets:
metadata
license: cc-by-4.0
task_categories:
- visual-question-answering
language:
- en
pretty_name: BedBench
size_categories:
- 1K<n<10K
configs:
- config_name: T1_uncolor
data_files:
- split: train
path: data/T1_uncolor/train-*.parquet
- config_name: T1_uncolor_axes
data_files:
- split: train
path: data/T1_uncolor_axes/train-*.parquet
- config_name: T1_color
data_files:
- split: train
path: data/T1_color/train-*.parquet
- config_name: T2_uncolor
data_files:
- split: train
path: data/T2_uncolor/train-*.parquet
- config_name: T2_uncolor_axes
data_files:
- split: train
path: data/T2_uncolor_axes/train-*.parquet
- config_name: T3_uncolor_axes
data_files:
- split: train
path: data/T3_uncolor_axes/train-*.parquet
- config_name: T3_color_axes
data_files:
- split: train
path: data/T3_color_axes/train-*.parquet
- config_name: T4_color
data_files:
- split: train
path: data/T4_color/train-*.parquet
- config_name: T4_color_axes
data_files:
- split: train
path: data/T4_color_axes/train-*.parquet
- config_name: T5_color
data_files:
- split: train
path: data/T5_color/train-*.parquet
- config_name: T5_color_axes
data_files:
- split: train
path: data/T5_color_axes/train-*.parquet
- config_name: T5_finetune
data_files:
- split: train
path: data/T5_finetune/train-*.parquet
- split: validation
path: data/T5_finetune/validation-*.parquet
- config_name: T6_uncolor
data_files:
- split: train
path: data/T6_uncolor/train-*.parquet
- config_name: T6_uncolor_axes
data_files:
- split: train
path: data/T6_uncolor_axes/train-*.parquet
- config_name: T7_color
data_files:
- split: train
path: data/T7_color/train-*.parquet
- config_name: T7_color_axes
data_files:
- split: train
path: data/T7_color_axes/train-*.parquet
- config_name: T8_uncolor
data_files:
- split: train
path: data/T8_uncolor/train-*.parquet
- config_name: T8_uncolor_axes
data_files:
- split: train
path: data/T8_uncolor_axes/train-*.parquet
- config_name: T9_uncolor
data_files:
- split: train
path: data/T9_uncolor/train-*.parquet
- config_name: T9_uncolor_axes
data_files:
- split: train
path: data/T9_uncolor_axes/train-*.parquet
- config_name: T10_color
data_files:
- split: train
path: data/T10_color/train-*.parquet
- config_name: T10_color_axes
data_files:
- split: train
path: data/T10_color_axes/train-*.parquet
- config_name: T11_invalid_uncolor
data_files:
- split: train
path: data/T11_invalid_uncolor/train-*.parquet
- config_name: T11_invalid_uncolor_axes
data_files:
- split: train
path: data/T11_invalid_uncolor_axes/train-*.parquet
- config_name: T11_invalid_color
data_files:
- split: train
path: data/T11_invalid_color/train-*.parquet
- config_name: T11_valid_uncolor
data_files:
- split: train
path: data/T11_valid_uncolor/train-*.parquet
- config_name: T11_valid_uncolor_axes
data_files:
- split: train
path: data/T11_valid_uncolor_axes/train-*.parquet
- config_name: T11_valid_color
data_files:
- split: train
path: data/T11_valid_color/train-*.parquet
- config_name: T12_uncolor
data_files:
- split: train
path: data/T12_uncolor/train-*.parquet
- config_name: T12_uncolor_axes
data_files:
- split: train
path: data/T12_uncolor_axes/train-*.parquet
- config_name: T12_color
data_files:
- split: train
path: data/T12_color/train-*.parquet
- config_name: T13_invalid_uncolor
data_files:
- split: train
path: data/T13_invalid_uncolor/train-*.parquet
- config_name: T13_invalid_color
data_files:
- split: train
path: data/T13_invalid_color/train-*.parquet
- config_name: T13_valid_uncolor
data_files:
- split: train
path: data/T13_valid_uncolor/train-*.parquet
- config_name: T13_valid_color
data_files:
- split: train
path: data/T13_valid_color/train-*.parquet
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.