Dataset Viewer
Auto-converted to Parquet Duplicate
id
int64
category
string
description
string
why_hard
string
image
image
candidate_labels
list
expected_label
string
model_winner
string
model_scores_json
string
is_correct
bool
model_name
string
error_type
string
1
counting
Image has exactly 7 red circles. Does model know the count?
Vision models notoriously struggle with counting beyond ~4 objects
[ "seven red circles", "five red circles", "three red circles" ]
seven red circles
seven red circles
{"seven red circles": 0.99755859375, "five red circles": 0.97607421875, "three red circles": 0.48046875}
true
google/siglip2-base-patch16-224
none
2
spatial_relations
Orange square is on top of blue circle. Tests preposition binding.
Contrastive models often ignore relational structure; both labels share the same nouns
[ "an orange square on top of a blue circle", "a blue circle on top of an orange square" ]
an orange square on top of a blue circle
an orange square on top of a blue circle
{"an orange square on top of a blue circle": 1.0, "a blue circle on top of an orange square": 1.0}
true
google/siglip2-base-patch16-224
none
3
negation
Empty white plate. Tests whether model understands 'no food'.
CLIP-style models are known to largely ignore negation in text
[ "a plate with no food on it", "a plate with food on it" ]
a plate with no food on it
a plate with no food on it
{"a plate with no food on it": 0.05224609375, "a plate with food on it": 0.002132415771484375}
true
google/siglip2-base-patch16-224
none
4
orientation
Triangle pointing downward (upside-down). Tests orientation awareness.
Orientation is a subtle visual feature often ignored in web-crawled training data
[ "an upside-down triangle", "a triangle pointing up" ]
an upside-down triangle
a triangle pointing up
{"an upside-down triangle": 0.81298828125, "a triangle pointing up": 0.94482421875}
false
google/siglip2-base-patch16-224
orientation
5
color_identification
Teal/cyan rectangle on dark background. Tests precise color naming.
Boundary colors like teal (between blue and green) challenge color-text alignment
[ "a teal rectangle", "a green rectangle", "a blue rectangle" ]
a teal rectangle
a teal rectangle
{"a teal rectangle": 0.99755859375, "a green rectangle": 0.97900390625, "a blue rectangle": 0.93701171875}
true
google/siglip2-base-patch16-224
none
6
size_comparison
One large circle and one small circle. Tests relative size reasoning.
Relative size requires comparing objects within the scene, not just recognizing them
[ "a large circle and a small circle", "two circles of the same size" ]
a large circle and a small circle
a large circle and a small circle
{"a large circle and a small circle": 0.99072265625, "two circles of the same size": 0.98095703125}
true
google/siglip2-base-patch16-224
none
7
text_in_image
Image showing the word EXIT. Tests OCR-level vision-language binding.
Base vision-language models aren't trained for OCR; text reading is unreliable
[ "a sign that says EXIT", "a sign that says ENTER", "a sign that says STOP" ]
a sign that says EXIT
a sign that says EXIT
{"a sign that says EXIT": 0.0035648345947265625, "a sign that says ENTER": 0.0032482147216796875, "a sign that says STOP": 0.00257110595703125}
true
google/siglip2-base-patch16-224
none
8
shape_finegrain
A wide ellipse (clearly not a circle). Tests fine-grained shape discrimination.
Circle and ellipse are visually similar; both are 'oval' shapes in natural language
[ "a red ellipse", "a red circle" ]
a red ellipse
a red ellipse
{"a red ellipse": 0.9970703125, "a red circle": 0.79833984375}
true
google/siglip2-base-patch16-224
none
9
composition_vs_components
Three circles arranged in a triangle pattern. Part vs. whole.
Requires understanding both components AND their spatial arrangement as a gestalt
[ "three circles arranged in a triangle", "a triangle", "three separate circles" ]
three circles arranged in a triangle
three circles arranged in a triangle
{"three circles arranged in a triangle": 0.99072265625, "a triangle": 0.021942138671875, "three separate circles": 0.880859375}
true
google/siglip2-base-patch16-224
none
10
quantity_fullness
Container ~85% full of blue liquid. Tests quantity/fullness language.
Quantitative fullness requires calibrated visual-language grounding
[ "a nearly full container", "a half-empty container", "an empty container" ]
a nearly full container
a half-empty container
{"a nearly full container": 0.0017271041870117188, "a half-empty container": 0.0299835205078125, "an empty container": 0.010986328125}
false
google/siglip2-base-patch16-224
quantity_fullness
README.md exists but content is empty.
Downloads last month
5