Dataset Viewer
Auto-converted to Parquet Duplicate
sample_id
stringclasses
9 values
scene
stringclasses
9 values
view_a
stringclasses
5 values
view_b
stringclasses
8 values
option_a
stringclasses
9 values
option_b
stringclasses
9 values
task
stringclasses
1 value
correct_option
stringclasses
1 value
consistency_axis
stringclasses
9 values
difficulty
stringclasses
3 values
CVSC-0001
Two balls collide on table
Left camera view
Right camera view
Both views imply same post-collision motion
Views imply different outcomes
Choose the consistent interpretation
A
collision_state
1
CVSC-0002
Box slides and stops
Front view
Rear view
Both views show deceleration to stop
Rear view suggests later acceleration
Choose the consistent interpretation
A
motion_state
1
CVSC-0003
Cup placed on table
Eye-level view
Top-down view
Cup remains on table in both views
Top-down view implies disappearance
Choose the consistent interpretation
A
object_persistence
1
CVSC-0004
Person opens door inward
Side view
Opposite side view
Door opens inward in both views
Opposite view implies outward opening
Choose the consistent interpretation
A
constraint_direction
2
CVSC-0005
Ball rolls down ramp
Side view
Rear view
Both views show downhill acceleration
Rear view implies uphill motion
Choose the consistent interpretation
A
gravity_alignment
2
CVSC-0006
Block stack is stable
Front view
Oblique view
Both views show stable stack
Oblique view implies floating block
Choose the consistent interpretation
A
support_state
1
CVSC-0007
Car stops at red light
Street-side view
Overhead view
Both views show full stop
Overhead view implies rolling through
Choose the consistent interpretation
A
traffic_state
2
CVSC-0008
Water poured into cup
Side view
Top view
Both views imply same fill level
Top view implies higher level
Choose the consistent interpretation
A
volume_state
2
CVSC-0009
Ball bounces once
Side view
Angled view
Both views imply same bounce height
Angled view implies higher rebound
Choose the consistent interpretation
A
energy_state
3

What this is

This dataset tests abstract state consistency across viewpoints.

Same event.
Different camera.
One state.

What it measures

• Whether abstraction survives camera change
• Whether state ties to structure, not pixels

A model that binds state to view will fail.

Expected model output

Return only:

A
or
B

Scoring

Use scorer.py.

Metrics:

• accuracy
• parse_rate

Suggested prompt

You are checking whether two viewpoints describe the same world state.

Scene: {scene}

View A: {view_a}

View B: {view_b}

A: {option_a} B: {option_b}

Choose the consistent interpretation. Answer with only A or B.

Trinity placement

This is Dataset I of the Abstract State Consistency Trinity.

Next datasets

• Modality State Consistency
• Description Granularity Consistency

Citation

ClarusC64. "Clarus Viewpoint State Consistency v0.1", 2026.

Downloads last month
19