SERGE (SEmantic Richness & Geometric Evaluation)

SERGE is the evaluation dataset released with the paper Semantic Richness or Geometric Reasoning? The Fragility of VLM's Visual Invariance. It contains the canonical (untransformed) images used to test whether VLMs judge object identity under rotation, scale, and identity transformations through geometric reasoning or through familiarity with semantically rich content.

• Paper: arXiv:2604.01848
• Project page: xthomasbu.github.io/visual_invariance
• Code: github.com/XThomasBU/semantic_richness

Dataset overview

Dataset	# Samples	Description
Omniglot	1,623	Handwritten characters across 50 scripts
Times New Roman	52	Printed English alphabet (A–Z, a–z)
Handwritten English	52	Handwritten English alphabet
PACS	800	4 domains × 200 images each

Together these span a semantic-richness spectrum used in the paper: symbolic sketches (Omniglot) → semantic sketches (Handwritten English, Times New Roman) → natural photos and art (PACS).

Directory layout

omniglot/images_all/{Script}/characterXX/*.png   # 50 official scripts + hand_english
times_new_roman/characterXX/image.png
pacs/{domain}_{label}_{hf_index}.png
manifest.json                 # per-script / per-domain counts

hand_english (the Handwritten English row above) is stored under omniglot/images_all/hand_english alongside the 50 official Omniglot scripts. pacs filenames encode metadata, e.g. art_painting_0_101.png → domain=art_painting, label=0 (dog), source index=101.

Loading the dataset

from huggingface_hub import snapshot_download
from datasets import load_dataset

local_dir = snapshot_download(repo_id="XThomasBU/SERGE", repo_type="dataset")

omniglot = load_dataset("imagefolder", data_dir=f"{local_dir}/omniglot/images_all", split="train")
times_new_roman = load_dataset("imagefolder", data_dir=f"{local_dir}/times_new_roman", split="train")
pacs = load_dataset("imagefolder", data_dir=f"{local_dir}/pacs", split="train")

Citation

If you use SERGE, please cite the paper:

@article{qiu2026semantic,
  title={Semantic Richness or Geometric Reasoning? The Fragility of VLM's Visual Invariance},
  author={Qiu, Jason and Meurer, Zachary and Thomas, Xavier and Ghadiyaram, Deepti},
  journal={arXiv preprint arXiv:2604.01848},
  year={2026}
}

and the original source datasets:

Brenden M. Lake, Ruslan Salakhutdinov, and Joshua B. Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332–1338, 2015. doi: 10.1126/science.aab3050. URL https://www.science.org/doi/abs/10.1126/science.aab3050.

Sujay Mann. Handwritten english characters and digits, 2024. URL https://www.kaggle.com/datasets/sujaymann/handwritten-english-characters-and-digits. Accessed: 2026-03-29.

Samuel Yu, Peter Wu, Paul Pu Liang, Ruslan Salakhutdinov, and Louis-Philippe Morency. Pacs: A dataset for physical audiovisual commonsense reasoning, 2022. URL https://arxiv.org/abs/2203.11130.

Stanley Morison and Victor Lardent. Times new roman. https://en.wikipedia.org/wiki/Times_New_Roman, 1932. Typeface originally commissioned for The Times newspaper. Accessed: 2026-03-30.