uc-training: yes/no questions about images (training data)
The training data 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). The evaluation suite is
LGirrbach/uc-benchmark, which uses the same layout and image ids.
5,671,272 questions on 352,960 images (52.3 GiB). This is training mix v6t_excl_hardneg_skills.
| split | images | questions |
|---|---|---|
train |
345,225 | 5,449,032 |
validation |
7,735 | 222,240 |
Train answers: 34.7% yes, 56.2% no, 9.1% uncertain.
Loading
from datasets import load_dataset
questions = load_dataset("LGirrbach/uc-training", "questions", split="train") # image_id, slice, question, answer, soft, type, boxes
images = load_dataset("LGirrbach/uc-training", "images", split="train") # image_id, image (PIL)
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 config has a train and a validation split. The train and validation images are disjoint, so the
questions of each split only refer to images of the same split.
| config | column | content |
|---|---|---|
questions |
image_id |
key into images (same split) |
slice |
the training slice the question comes from (table below) | |
question |
the question, or for VSR and the spatial slice the statement to verify | |
answer |
yes, no or uncertain (verifiers or human annotators disagree; usually dropped or down-weighted) |
|
soft |
soft label P(yes) in [0, 1]: the verifier MLLMs' confidence for MLLM-labelled items (attrgen, finesynth, hardneg, part of OpenImages), the share of yes answers for VQAv2, otherwise the hard label (0 / 1). For uncertain rows it is the only label |
|
type |
question type, e.g. attribute, count, spatial_depth, paco_part_color, excl_cars_make, hardneg (363 types; see stats.json) |
|
boxes |
grounding boxes for the two pointer slots, [[x0, y0, x1, y1], [x0, y0, x1, y1]] normalised to [0, 1] (x by width, y by height), or null. Spatial statements: [first-mentioned object, second-mentioned object]; part questions: [part, whole object] (absent parts: [object, object]); everything else repeats the queried object or region box. Present for 37% of train questions |
|
images |
image_id |
<image source>/<path in that source>, e.g. coco2017/train2017/000000200195.jpg |
image |
the image (original file bytes) |
Each image is stored once, however many questions refer to it. The image ids use the same scheme as
LGirrbach/uc-benchmark. No image of uc-benchmark appears here: all benchmark images, including the whole COCO
val2017 set, one held-out COCO train2017 shard and the TextOCR / TextVQA validation images, were removed before the
mix was built.
Slices
| slice | train images | train questions | yes | no | uncertain | val images | val questions |
|---|---|---|---|---|---|---|---|
attrgen |
54,922 | 1,582,569 | 23.7% | 54.9% | 21.3% | 4,984 | 145,685 |
human_yesno |
89,252 | 345,836 | 46.9% | 47.5% | 5.6% | 4,759 | 19,194 |
grounded |
33,537 | 664,728 | 38.7% | 61.3% | 0.0% | 121 | 2,644 |
spatial |
20,035 | 109,635 | 50.0% | 50.0% | 0.0% | 1,383 | 8,283 |
finesynth |
21,430 | 396,988 | 29.2% | 36.1% | 34.7% | 0 | 0 |
extdata |
223,354 | 628,690 | 50.2% | 49.8% | 0.0% | 2,208 | 6,197 |
exclusive |
62,238 | 464,557 | 23.2% | 76.8% | 0.0% | 594 | 4,386 |
hardneg |
67,972 | 305,266 | 50.0% | 50.0% | 0.0% | 0 | 0 |
count |
58,709 | 462,351 | 26.8% | 73.2% | 0.0% | 2,906 | 22,720 |
negation |
124,983 | 304,054 | 50.2% | 49.8% | 0.0% | 4,734 | 11,343 |
ocr |
21,447 | 184,358 | 39.8% | 60.2% | 0.0% | 214 | 1,788 |
An image can occur in several slices, so the image counts add up to more than the split total.
| slice | content |
|---|---|
attrgen |
Generated questions on COCO train2017 about attributes, scene and counts. An LLM proposes the questions, an MLLM selects them and two MLLMs verify the labels. Items where the verifiers disagree are kept as uncertain. |
human_yesno |
Human-labelled yes/no questions on COCO images: VQAv2 yes/no questions, GQA balanced yes/no questions and VSR statements. Each training question string is capped per label (diversity over quantity). |
grounded |
Object attributes, part attributes and part presence from PACO-LVIS and VAW ("Is the mug ceramic?", "Does the chair have an armrest?"). |
spatial |
Left/right, above/below and closer/farther statements generated from COCO boxes and Depth-Anything-V2 depth. Every surface form occurs with both labels. |
finesynth |
Region-grounded attribute labels on COCO: SAM3 regions, an MLLM attribute schema, programmatic checks and two-MLLM verification, rendered with six question templates. |
extdata |
Non-COCO questions built from structured annotations of OpenImages V7 (human-verified labels), Fashionpedia, iNat2021, WikiArt, CUB-200-2011, Stanford Cars, FGVC-Aircraft and DTD. Each question string is asked on a positive and a negative image. |
exclusive |
Mutually exclusive label families on the extdata train images (bird and iNat species; car make, model, body type and country; aircraft manufacturer, family and country; WikiArt artist, genre and style). Each group asks the true value and three sibling values (plus one never-true country for the country families) on the same image, so the model learns that only one of them can be true. |
hardneg |
Open-vocabulary hard negatives: concepts that a dual encoder (SigLIP2 / FG-CLIP2) ranks highly on an image but that two MLLMs (Gemma-4-26B, Qwen3-VL-32B) reject. Balanced per concept with verified positives (50% yes) and rendered with six templates. |
count |
Exact counting. Per (image, category): the true count n in 0..10 and the three nearest wrong counts in one template ("Are there exactly {k} {X} in the image?", "Is there exactly one {X} ...", "Are there no {X} ..."), half of the groups also with an "at least k" question (type count_atleast). COCO train2017 instance counts (all instances non-crowd and at least 0.1% of the image; n = 0 from absent categories of the same supercategory) and TallyQA train human counts, whose "How many ...?" questions an LLM (Qwen3.8-27B) turned into yes/no templates. |
negation |
Negated questions with flipped answers on the images of the other slices: "Is the car red?" (yes) becomes "Is the car not red?" (no). "Yes" always means the stated proposition holds; no contractions or inverted negative questions ("Isn't there ...?"). Existence questions are negated by rule, the others by an LLM (Qwen3-8B), and every LLM negation was accepted by a second LLM (Gemma-4-26B-A4B) as an exact negation. Questions with conjunctions, quantifiers or modal verbs (where "not" is ambiguous) are excluded. type is neg_<type of the source question>. |
ocr |
Reading text, on TextOCR / TextVQA train images: 'Does the image contain the text "{w}"?' (six templates) for legible annotated words at least 8 px high at 384 px resolution (yes), real words at edit distance 1 that are not in the image and words from other images (no); and TextVQA questions rewritten by an LLM (Qwen3.8-27B) into templates such as 'Is the brand of this camera "{answer}"?', asked with the human answer (yes), another word of the same image or an edit-distance-1 variant (no). Boxes are the word boxes. |
Notes
- Order: questions follow the mix file order, so all questions about one image are consecutive.
- Not exported: question groups (e.g. the exclusive-family group id used by the optional listwise loss).
Hence the count groups and the link between a
negationitem and its source question are not exported. - Image sources:
coco2017(104,518),openimages_v7(67,614),inat2021(49,924),wikiart(41,853),fashionpedia(41,777),textvqa(20,742),stanford_cars(8,142),fgvc_aircraft_2013b(6,664),cub_200_2011(5,992),dtd(3,757),visual_genome(1,977). OpenImages images are downscaled to a maximum side of 1024 px; all other images are the original files. The WikiArtimage_idis the parquet file and row ofhuggan/wikiart. - Licences: images and annotations keep the licences and terms of their original datasets. Generated questions and labels are for research use.
- Downloads last month
- 28