uc-benchmark: yes/no questions about images
The evaluation suite of the universal-classifier project: given an image and a yes/no question about it ("Is the
earring silver?", "Is the cat left of the dog?"), predict a calibrated P(yes). We report AUROC and ECE per benchmark.
696,936 questions on 31,067 images (4.9 GiB) from 16 benchmarks. All labels come from
existing human-labelled datasets, converted into a common binary format. For the skill benchmarks (counting,
negation, OCR) the questions were generated from templates or rewritten by an LLM; the labels are still human.
Loading
from datasets import load_dataset
questions = load_dataset("LGirrbach/uc-benchmark", "questions", split="test")
images = load_dataset("LGirrbach/uc-benchmark", "images", split="test")
idx = {iid: i for i, iid in enumerate(images["image_id"])}
q = questions[0]
img = images[idx[q["image_id"]]]["image"]
Hugging Face splits of one config must share their columns, so the two tables are two configs, questions and
images, each with a single test split.
| config |
column |
content |
questions |
image_id |
key into images |
|
benchmark |
benchmark name (table below) |
|
question |
the question, or for VSR and VALSE the statement / caption to verify |
|
answer |
yes or no |
images |
image_id |
<image source>/<path in that source>, e.g. coco2017/val2017/000000000139.jpg |
|
image |
the image (original file bytes) |
An image used by several benchmarks is stored once. The image ids match the image_source/image_path columns of
LGirrbach/uc-dataset (training data); none of these images appears in its train or validation split.
Benchmarks
Core
| benchmark |
images |
questions |
yes |
coco_exist |
4,929 |
393,879 |
3.5% |
vqav2 |
4,004 |
9,129 |
48.3% |
gqa |
1,496 |
7,581 |
50.9% |
vsr |
1,874 |
2,195 |
53.8% |
ovad |
500 |
15,023 |
50.7% |
vaw |
1,714 |
15,008 |
39.6% |
paco_part |
2,537 |
11,788 |
66.3% |
paco_attr |
629 |
15,010 |
50.0% |
valse |
6,069 |
15,404 |
50.0% |
| benchmark |
what it evaluates |
source paper |
coco_exist |
Object presence: "Is there a {category} in the image?" for all 80 COCO categories on 4.9k held-out COCO train2017 images, labelled by the COCO instance annotations. Only 3.5% yes (calibration under a skewed share of yes answers). |
COCO |
vqav2 |
Open-ended human questions about everyday scenes (objects, activities, attributes, counting, common sense). Kept: at least 8 of 10 human answers are yes/no and at least 70% agree. |
VQAv2 |
gqa |
Compositional, multi-step questions generated from human scene graphs (objects, relations, attributes, comparisons); balanced split. |
GQA |
vsr |
True/false statements about the spatial relation of two objects (66 relation types: left/right, front/behind, on, touching, facing, ...); official random-split test set. |
VSR |
ovad |
117 human-annotated attributes of uniquely referable objects on COCO val2017 (colour, material, pattern, texture, cleanliness, maturity, expression, colour quantity, position, ...). |
OVAD |
vaw |
"Is the {object} {attribute}?" on Visual Genome / COCO images: colour, material, shape, size, texture, state, activity. |
VAW |
paco_part |
Part presence on everyday objects: "Does the {object} have a {part}?" for optional parts. |
PACO |
paco_attr |
Colour, pattern/marking, material and transparency of objects and their parts (no same-hue colour negatives). |
PACO |
valse |
Caption vs. minimally changed foil over 11 linguistic phenomena: existence, plurals, counting (3), relations, action replacement, actant swap, coreference (2), FOIL-it. Only validated samples. |
VALSE, FOIL it! |
Additional existence benchmarks
| benchmark |
images |
questions |
yes |
lvis |
5,447 |
66,579 |
26.7% |
openimages_test |
5,000 |
33,305 |
60.8% |
| benchmark |
what it evaluates |
source paper |
lvis |
Existence over the 1,203 LVIS categories (long tail, including rare classes). LVIS is federated: per image, only the annotated categories (yes) and the verified-absent categories (no) are asked. Images: all COCO val2017 images in LVIS v1 val plus those of the held-out COCO train2017 shard (the other LVIS val images are COCO train2017 images used for training). |
LVIS |
openimages_test |
Existence over the 600 boxable OpenImages classes on 5,000 randomly sampled OpenImages V7 test images (seed 0, upright images only), using the human-verified image-level labels. Verified-absent labels were proposed by a classifier and rejected by annotators, so the negatives are hard. Images are downscaled to a maximum side of 1024 px. |
Open Images |
Skill benchmarks: counting, negation, OCR
| benchmark |
images |
questions |
yes |
count_coco |
2,002 |
24,816 |
9.1% |
count_tallyqa |
3,655 |
43,329 |
9.1% |
neg |
5,937 |
19,200 |
50.0% |
ocr_textocr |
2,970 |
20,060 |
38.4% |
ocr_textvqa |
1,703 |
4,630 |
45.1% |
| benchmark |
what it evaluates |
source paper |
count_coco |
Exact counting. One group per (image, category) with true count n in 0..10 (all instances non-crowd and at least 0.1% of the image; n = 0 is an absent category from the same supercategory as a present one). Every group asks all eleven values ("Are there no {X} in the image?", "Is there exactly one {X} in the image?", "Are there exactly {k} {X} in the image?"), exactly one of which is yes. At most 250 groups per true count. Held-out COCO images (val2017 and a held-out train2017 shard). |
COCO |
count_tallyqa |
Exact counting with human "How many ...?" questions, simple ("How many dogs are there?") and complex ("How many people are wearing hats?"). An LLM (Qwen3.8-27B) turned each question into yes/no templates, and all eleven values 0..10 are asked as for count_coco. TallyQA test on Visual Genome images. |
TallyQA |
neg |
Negation. Pairs of a question from coco_exist, lvis, vaw, ovad, paco_attr, vsr, gqa or vqav2 (1,200 pairs each, balanced yes/no) and its logical negation with the flipped answer ("Is the table short?" / "Is the table not short?"). "Yes" always means the stated proposition holds; contractions and inverted negative questions ("Isn't there ...?") are never used. Existence questions are negated by rule, the others by an LLM (Qwen3-8B); every LLM negation was accepted by a second LLM (Gemma-4-26B-A4B) as an exact negation. Questions where "not" would be ambiguous (conjunctions, quantifiers, modal verbs) are excluded. |
see the source benchmarks |
ocr_textocr |
Reading words: 'Does the image contain the text "{w}"?' for annotated legible words in the image (yes; at least 4 characters, or numbers with at least 2 digits), a real word at edit distance 1 that does not occur in the image (no; numbers: one digit changed), and words from other images (no). TextOCR validation images. |
TextOCR |
ocr_textvqa |
Reading in context: an LLM (Qwen3.8-27B) rewrote TextVQA questions into templates such as 'Is the brand of this camera "{answer}"?', asked with the majority human answer (yes; at least 4 of 10 annotators agree and the answer is annotated text in the image), another legible word of the same image that is no human answer (no), and a real word at edit distance 1 from the answer that does not occur in the image (no). About 2% of the templates are clumsy ('Is what is inside the building "..."?'). TextVQA validation images. |
TextVQA |
The TextOCR / TextVQA images are OpenImages train images; validation images that are also in the training data of
LGirrbach/uc-dataset were removed.
Notes
- Order: questions keep the benchmark file order. For
valse, each caption (yes) is immediately followed by its
foil (no) on the same image, so pairwise accuracy can be computed from consecutive rows. Likewise, every neg
pair is two consecutive rows (original, then negation), and every count_coco / count_tallyqa group is eleven
consecutive rows (k = 0, 1, ..., 10), so count accuracy (argmax over the group) can be computed from row order.
- Not exported: question types (e.g. VSR relation, OVAD attribute type, VALSE subset), soft labels, grounding
boxes.
- Shared images:
coco_exist, vqav2 and gqa share most images (one held-out COCO shard); paco_part and
paco_attr share 415 images; lvis shares images with ovad and coco_exist.
- Image sources: coco2017 (17,254), openimages_v7 (5,000), visual_genome (4,637), textvqa (2,970), swig (1,102), visdial_v1.0_val2018 (104). Images and annotations keep the licences and terms of their original datasets (see the
linked papers / dataset pages). This collection is for research evaluation.