uc-dataset: binary image–question supervision for a calibrated universal classifier
This dataset trains and evaluates a model that returns a calibrated probability P(yes | image, question) for
any yes/no question or statement about an image (for example "Is the earring silver?", "The cat is left of the dog.",
"Was this painted by Rembrandt?"). It combines human annotations from 20+ public datasets with labels from
MLLM pipelines and programmatic rules. The dataset holds text and labels only. Images are referenced by
(image_source, image_path) and must be downloaded from the original datasets (see
Getting the images).
Code: github.com/LGirrbach/universal-classifier (code/data_generation/export_hf.py builds this dataset from the
training mix default_xfer and the benchmarks in data/benchmarks/).
Columns
| column | type | meaning |
|---|---|---|
slice |
string | train / validation: the training slice the sample comes from (attrgen, xfer, human_yesno, spatial, grounded, finesynth, extdata). test: the benchmark name (vqav2, ovad@h1, valse, ...) |
image_source |
string | the dataset the image is taken from (coco2017, openimages_v7, inat2021, wikiart, ...) |
image_path |
string | how to identify the image in that dataset (relative path, file name or parquet row; see the table below) |
prompt |
string | the yes/no question or the statement to verify |
answer |
string | yes, no, or uncertain |
answer = "uncertain" marks items whose labellers disagree: the two MLLM verifiers in attrgen, xfer and
finesynth, or the 10 human answers in VQAv2 (yes share between 0.3 and 0.7). Our default training run leaves them out. To match it, filter with
answer != "uncertain". Every row is one (image, prompt) pair. An image appears in as many rows as it has
questions.
from datasets import load_dataset
ds = load_dataset("LGirrbach/uc-dataset", token=True) # private repo: needs an HF token with access
train = ds["train"].filter(lambda r: r["answer"] != "uncertain")
vsr = ds["test"].filter(lambda r: r["slice"] == "vsr")
Splits
| split | rows | images | content |
|---|---|---|---|
train |
5,652,545 | 323,972 | training mix default_xfer (all seven slices) |
validation |
274,337 | 7,507 | the validation part of the same slices (model selection, Platt calibration) |
test |
1,287,538 | 79,437 | all 37 benchmarks (21 official third-party benchmarks + 16 of our own validation and diagnostic sets) |
Disjointness. No image from train or validation appears in any test benchmark. The build removes every benchmark image from every slice, and it also keeps all 5,000 COCO val2017 images out of train and validation. Train and validation are disjoint by image.
Training slices (train / validation)
| slice | train rows (yes / no / uncertain) | train images | val rows | image source | labels |
|---|---|---|---|---|---|
attrgen |
2,899,117 (687,555 / 1,592,454 / 619,108) | 100,616 | 155,755 | COCO train2017 | An LLM proposes attribute, scene, count and existence questions per image; an MLLM selects them; two MLLMs verify (Qwen3.8-27B-FP8, Gemma-4-26B-A4B-it). uncertain = the verifiers disagree |
xfer |
612,947 (208,101 / 273,636 / 131,210) | 108,251 | 82,337 | COCO train2017 | cross-image transfer: attrgen question strings asked on other COCO images and verified by the same two MLLMs, to balance P(yes | question) |
human_yesno |
344,085 (161,129 / 163,763 / 19,193) | 89,039 | 19,128 | COCO train2017 | human: VQAv2 yes/no questions (uncertain = annotators split), GQA balanced yes/no, VSR train (statements) |
spatial |
109,552 (54,782 / 54,770 / 0) | 20,031 | 8,282 | COCO train2017 | programmatic left/right, above/below and closer/farther questions between COCO objects (COCO boxes + Depth-Anything-V2) |
grounded |
661,301 (254,824 / 406,477 / 0) | 33,508 | 2,638 | COCO train2017 (PACO-LVIS, VAW via VG→COCO) | human object and part attributes from PACO-LVIS and VAW |
finesynth |
396,895 (115,701 / 143,481 / 137,713) | 21,430 | – | COCO train2017 | FineSynth region attributes (SAM 3 regions + attribute schema + two-MLLM verification), with diverse question templates and referring expressions |
extdata |
628,648 (315,411 / 313,237 / 0) | 223,349 | 6,197 | OpenImages V7, Fashionpedia, iNat2021, WikiArt, CUB, Stanford Cars, FGVC-Aircraft, DTD | human labels and dataset metadata. Most questions come as a (positive image, negative image) pair with the identical string, so P(yes | question) = 0.5. OpenImages negatives and region attributes are confirmed by both MLLMs |
Validation = all questions on the COCO images of attrgen shard 011 and of VSR dev, PACO-LVIS / VAW val, and a 1 % hash split of every external (non-COCO) source. FineSynth has no validation part.
Test benchmarks (slice = benchmark name)
Official (third-party, human labels):
| benchmark | rows | images | image_source | content |
|---|---|---|---|---|
coco_exist |
393,879 | 4,929 | coco2017 | existence of each of the 80 COCO categories per image (COCO instance annotations) on held-out COCO train2017 images |
vqav2 |
10,127 | 4,140 | coco2017 | VQAv2 yes/no questions on held-out COCO images (uncertain = annotators split) |
gqa |
7,581 | 1,496 | coco2017 | GQA balanced yes/no questions on held-out COCO images |
vsr |
2,195 | 1,874 | coco2017 | VSR (random split) test statements |
ovad |
15,023 | 500 | coco2017 (val2017) | OVAD object attributes, human positive and negative labels |
paco_attr |
15,010 | 629 | coco2017 | PACO-LVIS test object and part attributes |
paco_part |
11,788 | 2,537 | coco2017 | PACO-LVIS test: presence of optional object parts |
vaw |
15,008 | 1,714 | coco2017 | VAW test attributes (on VG images that have a COCO id) |
valse |
15,404 | 6,069 | coco2017, visual_genome, swig, visdial_v1.0_val2018 | VALSE caption/foil pairs (11 phenomena, validated samples only): caption = yes, foil = no |
cub |
18,126 | 5,786 | cub_200_2011 | CUB-200 species and part attributes (official test split) |
cars |
30,028 | 8,039 | stanford_cars | Stanford Cars model, make and body type (cars_test) |
aircraft |
13,090 | 3,332 | fgvc_aircraft_2013b | FGVC-Aircraft variant, family and manufacturer (test split) |
inat |
23,908 | 10,000 | inat2021 | iNat2021 species, one val image per species |
dtd |
4,438 | 1,879 | dtd | DTD textures |
wikiart |
6,712 | 3,877 | wikiart | WikiArt artist, style and genre (8 % hash split) |
fashionpedia |
10,692 | 1,156 | fashionpedia | Fashionpedia garment attributes (val/test images) |
openimages |
69,419 | 27,539 | openimages_v7 | OpenImages V7 validation: human-label items only (existence, relations, attributes) |
ovad@h1..3, vaw@h1..3, paco_attr@h1..3, paco_part@h1..3 |
15,023 / 14,595 / 14,324 / 11,589 each | as above | as above | robustness: the same labels rephrased with three question templates that never occur in training (h1: "Would you say that …?", h2: "In this photo, …?", h3: "Check whether …") |
Validation and diagnostics (our own MLLM or programmatic labels, not official):
| benchmark | rows | images | content |
|---|---|---|---|
attrgen |
142,408 | 4,929 | attrgen pipeline on held-out COCO shard 012 (includes uncertain) |
attrgen_balanced |
64,762 | 4,929 | the same images, questions rebalanced per string with cross-image transfer labels |
finesynth, finesynth@h1..3 |
57,775 each | 4,607 | FineSynth region attributes on held-out images, plus three held-out templates |
spatial |
6,000 | 1,046 | programmatic spatial relations (COCO boxes + Depth-Anything-V2) on held-out images |
openimages_mllm |
14,247 | 6,771 | OpenImages V7 validation: MLLM-labelled region and attribute questions |
Getting the images
Download each source below and set a root directory per image_source; the image is then
<root>/<image_path> (WikiArt is the one exception: see below). Sizes are approximate download sizes.
| image_source | image_path format |
used in | how to get it |
|---|---|---|---|
coco2017 |
train2017/000000xxxxxx.jpg, val2017/000000xxxxxx.jpg |
train, val, test | http://images.cocodataset.org/zips/train2017.zip (18 GB), http://images.cocodataset.org/zips/val2017.zip (1 GB). Questions from VQAv2, GQA, VSR, VAW, PACO and VALSE (COCO 2014 / Visual7W) were mapped to COCO 2017 file names; the 12-digit id is the COCO image id. |
openimages_v7 |
train/<ImageID>.jpg, validation/<ImageID>.jpg |
train, val, test | https://s3.amazonaws.com/open-images-dataset/{train,validation}/<ImageID>.jpg, one image per file (or the official downloader, https://storage.googleapis.com/openimages/web/download_v7.html). We use about 67k train and 34k validation images, only those with Rotation == 0, downscaled to a maximum side of 1024 px. The questions do not depend on resolution. |
inat2021 |
train/<class dir>/<uuid>.jpg, val/<class dir>/<uuid>.jpg |
train, val, test | full iNat2021 train (not train_mini) and val: https://ml-inat-competition-datasets.s3.amazonaws.com/2021/train.tar.gz (224 GB), https://ml-inat-competition-datasets.s3.amazonaws.com/2021/val.tar.gz (8 GB); see https://github.com/visipedia/inat_comp/tree/master/2021 |
cub_200_2011 |
images/<class dir>/<file>.jpg |
train, val, test | CUB_200_2011.tgz from https://data.caltech.edu/records/65de6-vp158 (paths relative to CUB_200_2011/) |
stanford_cars |
cars_train/<5 digits>.jpg, cars_test/<5 digits>.jpg |
train, val, test | The original Stanford site is offline. Use a mirror of cars_train / cars_test with the devkit and cars_test_annos_withlabels.mat (for example the Kaggle dataset jessicali9530/stanford-cars-dataset). |
fgvc_aircraft_2013b |
data/images/<7 digits>.jpg |
train, val, test | https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/archives/fgvc-aircraft-2013b.tar.gz (paths relative to fgvc-aircraft-2013b/) |
dtd |
images/<class>/<file>.jpg |
train, val, test | https://www.robots.ox.ac.uk/~vgg/data/dtd/download/dtd-r1.0.1.tar.gz (paths relative to dtd/) |
fashionpedia |
train/<hash>.jpg, test/<hash>.jpg |
train, val, test | https://s3.amazonaws.com/ifashionist-dataset/images/train2020.zip → train/, https://s3.amazonaws.com/ifashionist-dataset/images/val_test2020.zip → test/ |
wikiart |
data/train-000NN-of-00072.parquet:<row> |
train, val, test | HF dataset huggan/wikiart at revision d559852d2b232e0fcf195e775866964f0564f2b5: the image is in column image of that parquet file at that row (0-based). Because the files are read in sorted order, the row also equals the global index in load_dataset("huggan/wikiart", split="train", revision=...). We stored the images downscaled to a 1024 px maximum side. |
visual_genome |
<VG image id>.jpg |
test (valse) |
Visual Genome v1.2 images: https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip and https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip (the file is in VG_100K/ or VG_100K_2/). VALSE Visual7W images that have a COCO id are listed under coco2017 instead. |
swig |
<verb>_<n>.jpg |
test (valse) |
SWiG / imSitu images: https://swig-data-weights.s3.us-east-2.amazonaws.com/images_512.zip (13.6 GB; only 1.1k files are needed, which you can extract with HTTP range requests) |
visdial_v1.0_val2018 |
VisualDialog_val2018_<12 digits>.jpg |
test (valse) |
VisDial v1.0 validation images (VisualDialog_val2018.zip), https://visualdialog.org/data |
Example resolver:
import io
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download
from PIL import Image
ROOTS = {"coco2017": "/data/coco/images", "openimages_v7": "/data/openimages/images", "inat2021": "/data/inat2021",
"cub_200_2011": "/data/CUB_200_2011", "stanford_cars": "/data/stanford_cars",
"fgvc_aircraft_2013b": "/data/fgvc-aircraft-2013b", "dtd": "/data/dtd", "fashionpedia": "/data/fashionpedia",
"visual_genome": "/data/VisualGenome/images", "swig": "/data/swig/images_512",
"visdial_v1.0_val2018": "/data/visdial/VisualDialog_val2018"}
WIKIART_REV = "d559852d2b232e0fcf195e775866964f0564f2b5"
def load_image(row):
if row["image_source"] == "wikiart":
file, r = row["image_path"].split(":")
local = hf_hub_download("huggan/wikiart", file, repo_type="dataset", revision=WIKIART_REV)
img = pq.read_table(local, columns=["image"]).slice(int(r), 1).column("image")[0].as_py()
return Image.open(io.BytesIO(img["bytes"])).convert("RGB")
return Image.open(f"{ROOTS[row['image_source']]}/{row['image_path']}").convert("RGB")
Rebuilding the labels from raw annotations
You only need the images to use this dataset. To rebuild the slices and benchmarks with the project code, you also
need these raw annotations (stored in the project under data/external/ unless noted otherwise):
| dataset | files | source |
|---|---|---|
| COCO 2017 | annotations_trainval2017.zip (instances; coco_exist, spatial, attrgen object lists) |
http://images.cocodataset.org/annotations/annotations_trainval2017.zip |
| VQAv2 | v2_{OpenEnded_mscoco_,mscoco_}{train,val}2014_{questions,annotations}.json |
https://s3.amazonaws.com/cvmlp/vqa/mscoco/vqa/v2_{Questions,Annotations}_{Train,Val}_mscoco.zip, https://visualqa.org/download.html |
| GQA | {train,val,testdev,test,challenge}_balanced_questions.json |
https://downloads.cs.stanford.edu/nlp/data/gqa/questions1.2.zip |
| Visual Genome | image_data.json (VG id → COCO id, for GQA, VAW and VALSE) |
https://homes.cs.washington.edu/~ranjay/visualgenome/data/dataset/image_data.json.zip |
| VSR | {train,dev,test}.jsonl (random split) |
https://huggingface.co/datasets/cambridgeltl/vsr_random/resolve/main/{train,dev,test}.jsonl |
| OVAD | ovad2000.json |
https://lmb.informatik.uni-freiburg.de/resources/datasets/ovad.zip, https://github.com/OVAD-Benchmark/ovad-benchmark-code |
| PACO-LVIS | paco_lvis_v1_{train,val,test}.json |
https://dl.fbaipublicfiles.com/paco/annotations/paco_lvis_v1.zip |
| VAW | train_part1.json, train_part2.json, val.json, test.json, attribute_index.json, attribute_types.json, attribute_parent_types.json |
https://raw.githubusercontent.com/adobe-research/vaw_dataset/main/data/<file> |
| VALSE | data/*.json (11 subsets) |
https://github.com/Heidelberg-NLP/VALSE |
| OpenImages V7 | class descriptions, attributes and relationships descriptions, VRD train/val, bbox train/val, human image labels train/val, train-images-boxable-with-rotation.csv |
https://storage.googleapis.com/openimages/{v5,v6,v7,2018_04}/... (exact list: code/slurm/download_openimages.sh) |
| Fashionpedia | instances_attributes_{train,val}2020.json |
https://s3.amazonaws.com/ifashionist-dataset/annotations/... (code/slurm/download_misc.sh) |
| WikiArt | artist / style / genre labels | in the huggan/wikiart parquet files (code/data_generation/extdata/wikiart_extract.py) |
| CUB / Cars / Aircraft / DTD / iNat2021 | class labels, CUB attributes and parts, Cars devkit .mat files, Aircraft variant/family/manufacturer lists, iNat val.json |
in the archives listed above |
The MLLM-derived slices (attrgen, xfer, finesynth, OpenImages verification) were made with Qwen3.8-27B-FP8 and
Gemma-4-26B-A4B-it served by vLLM (about 75 H100-hours for attrgen). Their outputs appear here only through
prompt / answer.
Not included
- Soft labels (verifier probabilities for
uncertainitems), question type and group ids, and the grounding boxes (one or two normalised boxes per region question) that the project uses for an auxiliary grounding loss. They remain in the project's JSONL files. - Images (see above).
Licence and attribution
The questions and answers are derived from the datasets listed above. Each keeps its own licence and terms of use (COCO: CC BY 4.0 annotations, Flickr image terms; OpenImages: CC BY 4.0 annotations, CC BY 2.0 images; VQAv2, GQA, VSR, OVAD, PACO, VAW, VALSE, Fashionpedia, iNat2021, CUB, Stanford Cars, FGVC-Aircraft, DTD, WikiArt, Visual Genome, SWiG and VisDial each under their own terms). Please cite the original datasets when you use the corresponding slices or benchmarks.
- Downloads last month
- 30