PoVisLE

PoVisLE (Polish Vision-Language Evaluation) is a Polish vision-language benchmark for evaluating culturally grounded multimodal understanding. The full benchmark contains 1,117 images and 2,366 manually annotated VQA pairs. This public release contains only the validation split, with 406 VQA pairs. A single example of the dataset open-ended task is presented below:

Dataset Structure

The dataset is released as three task configurations:

Configuration	Split	VQA pairs	Task
mcq	validation	212	multiple-choice
yn	validation	154	yes/no
open	validation	40	open-ended
Total		406

Questions are written in Polish and are organized into seven main categories.

Categories

The dataset is organized into the following main categories:

• Art and Entertainment: Polish and Poland-related artistic, media, and cultural references, including architecture, film, literature, music, paintings, sculpture, sport, and media.
• Culture and Tradition: shared customs, practices, symbols, cuisine, religion, traditions, pop culture, and regional or ethnic cultural variation.
• Geography and Nature: Polish physical, natural, urban, infrastructural, and socio-political spaces, including landscapes, landmarks, regions, and administrative entities.
• History and Society: historical and contemporary Polish social context, covering periods from the Middle Ages through World War II, post-war history, and current affairs.
• Language: visually grounded Polish linguistic phenomena, such as colloquial speech, slang, dialects, regionalisms, grammar, orthography, phraseology, rhetorical figures, and semantics.
• Image Understanding: direct recognition and interpretation of visual content in Polish or Poland-related contexts.
• Visual Reasoning: reasoning over relationships, context, or inferred information within an image.

Data Fields

Common fields across all configurations:

• id: unique sample identifier.
• image_id: image identifier.
• question_index: question index within a given image.
• image: input image.
• image_source: image provenance label, one of wiki, own, or other.
• image_url: source URL for images from external sources, when available.
• category: top-level taxonomy category.
• subcategory: fine-grained taxonomy category.
• taxonomy_path: full taxonomy path in the category > subcategory format.
• question: question in Polish.
• answer: gold answer.

Additional fields for mcq:

• A-H: answer options. Not every example uses all option fields.
• answer: the correct option label.

For yn, answers are Polish yes/no labels. For open, answers are short free-form gold responses.

Dataset Creation

PoVisLE was created through manual, template-free annotation. Annotators selected or reviewed images from Wikimedia Commons, other permissively available public sources, and personal collections contributed for research use. Each image was paired with one or more Polish VQA prompts and labeled with a task type, category, and subcategory.

The dataset construction also included a Wikimedia-based augmentation stage to increase visual diversity and reduce selection bias. Candidate images were reviewed by annotators, and visually similar replacements were used only when the original question remained answerable from the new image.

Quality assurance included cross-validation by a second annotator, metadata and license checks, regular team discussion, and supervision by an expert annotator. The annotation guidelines required questions to be visually grounded, unambiguous, linguistically natural, and suitable for deterministic evaluation.

Intended Use

PoVisLE is intended for research evaluation of vision-language models on Polish culturally and linguistically grounded visual question answering. It can be used to compare models, analyze performance across cultural and linguistic categories, and test whether model answers depend on both the image and the Polish question.

The benchmark is focused on a single, well-defined Polish context and supports controlled, fine-grained assessment of culturally situated vision-language understanding.

Ethical Considerations

The benchmark focuses on culturally situated Polish content and may reflect annotator perspectives and coverage choices. Some questions may privilege culturally prototypical answers over alternative but valid interpretations, so results should be interpreted within the dataset's scope.

License

The text annotations in this dataset are released under the CC BY-SA license. The images included in the dataset come from a combination of sources: images obtained from Wikipedia and other external sources retain the licenses assigned by their respective providers, while images created specifically for this dataset are released under the CC BY-SA license.

Users are responsible for complying with the applicable license terms for each individual image in addition to the license governing the dataset annotations.

Quickstart

To load the dataset, you can use the datasets library as follows:

from datasets import load_dataset

mcq = load_dataset("NASK-PIB/PoVisLE", "mcq", split="validation")
yn = load_dataset("NASK-PIB/PoVisLE", "yn", split="validation")
open_ended = load_dataset("NASK-PIB/PoVisLE", "open", split="validation")

example = mcq[0]
print(example["question"])
print({label: example[label] for label in "ABCDEFGH" if example.get(label)})
print(example["answer"])

Evaluation code is available at NASK-NLP/PoVisLE. A minimal evaluation run can be launched as:

python -m povisle.evaluate \
  --model-config configs/vllm/qwen3_5_9b_instruct.yml \
  --dataset-id NASK-PIB/PoVisLE \
  --split validation \
  --tasks all

Acknowledgement

This work was supported by the Polish Ministry of Digital Affairs (subsidy no. 4/WII/DBI/2026). The computational resources were provided by the Polish high-performance computing infrastructure PLGrid (HPC Center: ACK Cyfronet AGH) under computational grant no. PLG/2026/019138.

Citation

@article{kolos2026povisle,
  title   = {Jako Tako or Fluent? Presenting PoVisLE: A Polish Vision-Language Evaluation},
  author  = {Ko{\l}os, Anna and Statkiewicz, Grzegorz and Seweryn, Karolina and Kowol, Katarzyna and Piosek, Karolina and Kusa, Wojciech},
  journal = {arXiv preprint},
  year    = {2026}
}