Datasetpublic

KETI-NLP/K-prism

K-Prism Korean diagnostic benchmark data for evaluating hallucination in vision-language models. Evaluation code and protocol documentation are available at alsgur0720/K-Prism. Files File Contents Text_track.json 504 text-track questions Image_track.json 498 image-track questions images/ 165 original images referenced by the image track The release contains 1,002 questions and approximately 303 MB of annotations and images. Keep both JSON files… See the full description on the dataset page: https://huggingface.co/datasets/KETI-NLP/K-prism.

sourceHugging Faceupdated 6d agoView on Hugging Face

0likes192downloads

Dataset Card

K-Prism

Korean diagnostic benchmark data for evaluating hallucination in vision-language models. Evaluation code and protocol documentation are available at alsgur0720/K-Prism.

Files

File	Contents
`Text_track.json`	504 text-track questions
`Image_track.json`	498 image-track questions
`images/`	165 original images referenced by the image track

The release contains 1,002 questions and approximately 303 MB of annotations and images. Keep both JSON files and the images/ directory at the dataset repository root. Image paths are relative to Image_track.json (for example, images/example.png). Evaluation perturbations are generated by the evaluation code; they are not additional image files in this release.

Annotation fields

Both tracks contain id, image, question, passage, options, and explanation. Text annotations have a shared answer field. Image annotations have answer_a for the original-image condition and answer_b for the perturbed condition. Answer indices are one-based. Preserve the option ordering.

The text and image tracks are separate configurations because their answer fields differ. Both are exposed as evaluation (test) splits. The image column stores a relative path string; this raw JSON release does not automatically decode it into an image in the Dataset Viewer.

Download for evaluation

Download the complete repository using huggingface_hub.snapshot_download with repo_type="dataset" and the repository ID displayed on this page. Use the returned directory as the data root. From the evaluation code repository, run:

bash

python -m kprism validate \
  --text-data /path/to/download/Text_track.json \
  --image-data /path/to/download/Image_track.json

See the code repository for the full evaluation procedure and interpretation of intact versus perturbed conditions. The annotations include gold answers and explanations intended for evaluation.