README.md

metadata

configs:
  - config_name: all
    default: true
    data_files:
      - split: train
        path: data/all/train*.parquet
      - split: test
        path: data/all/test*.parquet
  - config_name: qrr
    data_files:
      - split: train
        path: data/qrr/train*.parquet
      - split: test
        path: data/qrr/test*.parquet
  - config_name: trr
    data_files:
      - split: train
        path: data/trr/train*.parquet
      - split: test
        path: data/trr/test*.parquet
  - config_name: fdr
    data_files:
      - split: train
        path: data/fdr/train*.parquet
      - split: test
        path: data/fdr/test*.parquet
task_categories:
  - visual-question-answering
language:
  - en
license: mit
tags:
  - spatial-reasoning
  - vlm-benchmark
  - ordinal-relations
  - 3d-scenes
  - multi-view
size_categories:
  - 100K<n<1M

ORDINARY-BENCH Dataset

A benchmark dataset for evaluating Vision-Language Models (VLMs) on ordinal spatial reasoning in 3D scenes.

Source code & evaluation pipeline: GitHub - tasd12-ty/ordinary-bench-core

Overview

Scenes	700 synthetic 3D scenes (Blender, CLEVR-style)
Complexity	7 levels: 4 to 10 objects per scene (100 each)
Questions	332,857 total across 3 reasoning types
Images	480 x 320 PNG, single-view (embedded in dataset)
Multi-view	4 camera angles per scene (available in source repo)

Question Types

QRR (Quantitative Relation Reasoning) -- 130,557 questions

Compare 3D distances between object pairs. Two variants:

• Disjoint: Is dist(A,B) less than, approximately equal to, or greater than dist(C,D)?
• Shared anchor: From anchor A, is dist(A,B) less/equal/greater than dist(A,C)?
• Answer format: <, ~=, or >

TRR (Ternary Relation Reasoning) -- 197,400 questions

Clock-face direction reasoning:

• Standing at object ref1, facing toward object ref2 (12 o'clock direction)
• What clock hour (1-12) is the target object at?
• Answer format: integer 1-12

FDR (Full Distance Ranking) -- 4,900 questions

Given an anchor object, rank all other objects by 3D distance, nearest to farthest.

• Answer format: ordered JSON array of object IDs, e.g., ["obj_2", "obj_1", "obj_3"]

Quick Start

from datasets import load_dataset

# Load QRR questions (test split)
ds = load_dataset("TYTSTQ/ordinary-bench", "qrr", split="test")

sample = ds[0]
sample["image"]                # PIL Image (480x320)
sample["question_text"]        # "Compare the distance between obj_0 and obj_1 vs ..."
sample["qrr_gt_comparator"]   # Ground truth: "<", "~=", or ">"

# Load all question types
ds_all = load_dataset("TYTSTQ/ordinary-bench", split="test")

# Load by specific type
ds_trr = load_dataset("TYTSTQ/ordinary-bench", "trr", split="test")
ds_fdr = load_dataset("TYTSTQ/ordinary-bench", "fdr", split="test")

Configs

Config	Description	Questions
all (default)	All 3 question types	332,857
qrr	Distance comparison only	130,557
trr	Clock direction only	197,400
fdr	Distance ranking only	4,900

Data Splits

Split	Scenes per complexity	Total scenes	Total questions
train	80	560	266,261
test	20	140	66,596

Column Schema

Common columns (all configs)

Column	Type	Description
scene_id	string	Scene identifier, e.g., n04_000080
n_objects	int	Number of objects in scene (4-10)
split	string	Complexity split: n04 through n10
image	Image	Rendered scene image (480x320 PNG)
objects	string	JSON array: [{"id": "obj_0", "desc": "large brown rubber sphere"}, ...]
question_type	string	qrr, trr, or fdr
qid	string	Question ID, e.g., qrr_0001
question_text	string	Natural language question
scene_metadata	string	Full scene JSON (3D coordinates, camera parameters, etc.)

QRR-specific columns

Column	Type	Description
qrr_variant	string	disjoint or shared_anchor
qrr_pair1	string	JSON: ["obj_0", "obj_1"]
qrr_pair2	string	JSON: ["obj_2", "obj_3"]
qrr_metric	string	Distance metric, e.g., dist3D
qrr_gt_comparator	string	Ground truth: <, ~=, or >

TRR-specific columns

Column	Type	Description
trr_target	string	Target object ID
trr_ref1	string	Standing position object
trr_ref2	string	12 o'clock facing direction object
trr_gt_hour	int	Ground truth clock hour (1-12)
trr_gt_quadrant	int	Ground truth quadrant (1-4)
trr_gt_angle_deg	float	Ground truth angle in degrees

FDR-specific columns

Column	Type	Description
fdr_anchor	string	Anchor object ID
fdr_n_ranked	int	Number of objects to rank
fdr_gt_ranking	string	JSON: ["obj_2", "obj_1", "obj_3"] (nearest to farthest)
fdr_gt_distances	string	JSON: [3.006, 3.553, 3.882]
fdr_gt_tie_groups	string	JSON: [["obj_2"], ["obj_1", "obj_3"]]

Scoring Criteria

Type	Metric	Description
QRR	Accuracy	Exact comparator match (<, ~=, >)
TRR	Hour accuracy	Exact clock hour match
TRR	Quadrant accuracy	Correct quadrant (1/4 of clock face)
TRR	Adjacent accuracy	Within +/-1 hour of ground truth
FDR	Exact accuracy	Full ranking match (respecting tie groups)
FDR	Kendall tau	Rank correlation coefficient [-1, 1]
FDR	Pairwise accuracy	Fraction of correct pairwise orderings
FDR	Top-1 accuracy	Nearest object correctly identified

Prompt Templates

System prompts for VLM evaluation are included in prompts/system_prompts.json. They instruct VLMs to respond with a JSON array of {"qid": "...", "answer": ...} objects.

Source Code

The full evaluation pipeline, scene generation code, and reconstruction tools are available at:

github.com/tasd12-ty/ordinary-bench-core

License

MIT