Dataset Card for VisLingA

Dataset Summary

VisLingA (Visual Lingual Analogies) is a dataset for testing whether various analogies hold both visually and linguistically. An analogy is a relationship like "Paris is to France as London is to the UK", or more generally A:B::C:D (A is to B as C is to D). VisLingA is based on the analogies dataset of Mikolov et al. (2013), taking a subset of the analogies and concepts therein and adding images representing those concepts. VisLingA was introduced in Persistent Representational Misalignment Induces Vision-Language Model Failures.

VisLingA was compiled by scraping about 5-10 images per concept from Wikimedia Commons, only taking images in the public domain or with a CC0/CC-BY license. For per-image attribution, see ATTRIBUTION.md.

Usage

# pip install -q datasets
from datasets import load_dataset
images = load_dataset("YGideoni/VisLingA", "images", split="test")
analogies = load_dataset("YGideoni/VisLingA", "analogies", split="test")

An embedding has an analogy A:B::C:D if D is the nearest concept to B - A + C out of all the concepts, with A, B and C excluded from the candidates. For example, evaluating a text model:

# pip install -q datasets sentence-transformers
import numpy as np
from datasets import load_dataset
from sentence_transformers import SentenceTransformer

analogies = load_dataset("YGideoni/VisLingA", "analogies", split="test")
concepts = sorted(set(analogies["word1"]) | set(analogies["word2"]) | set(analogies["word3"]) | set(analogies["word4"]))
idx = {c: i for i, c in enumerate(concepts)}

# text: embed each concept word (for images, embed each image of a concept and average)
model = SentenceTransformer("all-MiniLM-L6-v2")
emb = model.encode(concepts, normalize_embeddings=True)

a, b, c, d = (np.array([idx[w] for w in analogies[f"word{k}"]]) for k in range(1, 5))
target = emb[b] - emb[a] + emb[c]
target /= np.linalg.norm(target, axis=1, keepdims=True)
sims = target @ emb.T
rows = np.arange(len(d))
sim_d = sims[rows, d]
for k in (a, b, c):  # A, B and C are not candidates
    sims[rows, k] = -np.inf
ranks = 1 + (sims > sim_d[:, None]).sum(axis=1)
print(f"top-10 accuracy: {np.mean(ranks <= 10):.3f}, mean rank: {ranks.mean():.1f}")
# top-10 accuracy: 0.865, mean rank: 17.3

eval_vislinga.py in this repository runs the same evaluation for text and image models (sentence-transformers, Hugging Face image models and OpenCLIP), and also reports top-1 accuracy, median rank and per-category scores:

hf download YGideoni/VisLingA eval_vislinga.py --repo-type dataset --local-dir .
python eval_vislinga.py --model all-MiniLM-L6-v2 --modality text --out results/
python eval_vislinga.py --model openclip:ViT-B-16:laion2b_s34b_b88k --modality image --out results/

Dataset Structure

Data Instances

The dataset has two configs, each with a single test split: analogies (15,455 analogies over 746 concepts) and images (7,365 images, about 5-10 per concept). Every word in analogies is a concept in images.

An example from analogies:

{
  "analogy_id": "00000",
  "word1": "Athens",
  "word2": "Greece",
  "word3": "Baghdad",
  "word4": "Iraq",
  "category": "capital-common-countries"
}

An example from images:

{
  "image_id": "Athens/151201904",
  "concept": "Athens",
  "image": <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=875x1000>,
  "commons_pageid": 151201904,
  "commons_title": "Anafiotika - view of Athens with Saint George of the Rock Church's roof, 20240531 1204 9574.jpg",
  "source_page": "https://commons.wikimedia.org/wiki/File:Anafiotika_-_view_of_Athens_with_Saint_George_of_the_Rock_Church's_roof,_20240531_1204_9574.jpg",
  "author": "Jakub Hałun",
  "author_url": "https://commons.wikimedia.org/wiki/User:Jakubhal",
  "credit": "Own work",
  "license": "cc-by-4.0",
  "license_name": "CC BY 4.0",
  "license_url": "https://creativecommons.org/licenses/by/4.0/",
  "modifications": "Re-encoded as JPEG; Resized (longest side 1000 px)"
}

Data Fields

images:

• image_id: unique identifier, <concept>/<commons_pageid>
• concept: the concept the image represents
• image: the image (JPEG)
• commons_pageid: the image's Wikimedia Commons page id
• commons_title: the file's title on Wikimedia Commons
• source_page: URL of the file's Wikimedia Commons page
• author: the image's creator, as plain text ("Unknown" for some public-domain and CC0 images)
• author_url: link to the creator's page, if available
• credit: the credit line given on Wikimedia Commons, if any
• license: normalised license id: public-domain, cc0-1.0 or cc-by-<version>[-<jurisdiction>]
• license_name: the license as named on Wikimedia Commons
• license_url: link to the license text (null for public domain)
• modifications: changes made to the original file: "Converted to JPEG" or "Re-encoded as JPEG", plus "Resized (longest side 1000 px)" if it was downscaled

analogies:

• analogy_id: the analogy's index in the source analogy dataset
• word1, word2, word3, word4: A, B, C and D in the analogy A:B::C:D
• category: the analogy's category in the source dataset, e.g. capital-world or gram3-comparative

Licensing Information

All images are from Wikimedia Commons and keep their original licenses: 3,156 are in the public domain, 1,656 are CC0 1.0, and 2,553 are under a CC BY license (versions 1.0-4.0, some ported to a jurisdiction). Each image's license, license URL, author, source page and modifications are given in the images config, and ATTRIBUTION.md credits every CC BY image. Reusing a CC BY image requires attributing it as its license specifies. Images whose Commons license was ambiguous (e.g. listed differently in the metadata vs description) were excluded.

The analogies are derived from the Google analogy dataset of Mikolov et al. (2013), released with word2vec under the Apache License 2.0. The analogies config redistributes a modified version of word2vec's questions-words.txt (Copyright 2013 Google Inc.) under the Apache License 2.0: it keeps only the analogies whose four words all have images, and drops the plural, present-participle and adjective-to-adverb categories.

Everything else in the dataset, including the selection of concepts and images and the attribution metadata, is licensed under CC BY 4.0. The evaluation code, eval_vislinga.py, is licensed under the MIT License.

Citation Information

@article{gideoni2026persistent,
  title={Persistent Representational Misalignment Induces Vision-Language Model Failures},
  author={Gideoni, Yonatan and Gelberg, Yoav and Rudner, Tim GJ and Gal, Yarin},
  journal={arXiv preprint},
  year={2026}
}