PetQA: Benchmarking Veterinary Knowledge and Clinical Reasoning

PetQA is the first long-form question-answering dataset for veterinary medicine.

Paper · GitHub

PetQA provides text-only and multimodal question–answer pairs about dogs and cats, with reference answers written by veterinarians verified by the source platform.

Data collection and preprocessing

Data collection. We collected posts published from 2014 to 2024 on the pet medical consultation board of Naver Knowledge iN, a major online question-answering platform in South Korea.

Rule-based preprocessing. We retained questions about dogs and cats whose selected answers were written by verified veterinarians. We then removed duplicate questions and posts with corrupted images.

LLM-based preprocessing. We used GPT-4o-mini to filter out posts unrelated to pet medical consultation, unsupported or uninformative answers, personally identifiable information, promotional content, and noisy expressions while preserving essential clinical details.

After preprocessing, the final dataset contains 18,827 question–answer pairs. We split each modality into training, validation, and test sets.

Modality	Train	Validation	Test	Total
Text	6,076	2,000	2,000	10,076
Multimodal	4,751	2,000	2,000	8,751
Total	10,827	4,000	4,000	18,827

The test sets for benchmarking experiments are referred to as PetQA-Bench. PetQA-Bench is annotated under expert-reviewed guidelines and used to benchmark models' veterinary knowledge and clinical reasoning. The training and validation sets are used for supervised fine-tuning.

PetQA-Bench annotation

Each PetQA-Bench sample was annotated along two dimensions: question type and clinical conditions. These labels were designed to assess model performance across different clinical scenarios and determine whether model responses contain key clinical conditions.

To assess data quality, a veterinary expert with a Ph.D. reviewed 100 randomly sampled question–answer pairs using two criteria: relevance and correctness. Relevance assessed whether each answer appropriately addressed the question, while correctness assessed its factual accuracy. The review found that 99% of the sampled answers were relevant and 85% were factually correct. These results support the overall reliability of the dataset.

Multilingual extension

To broaden its global applicability, we used GPT-4o-mini to translate the original Korean benchmark into five additional languages: English, Chinese, German, Arabic, and Indonesian.

Dataset structure

data/
├── TEXT_ONLY/
│   ├── train.json
│   ├── validation.json
│   ├── test.json
│   ├── test_translated_en.json
│   ├── test_translated_zh.json
│   ├── test_translated_de.json
│   ├── test_translated_ar.json
│   └── test_translated_id.json
└── TEXT_w_IMAGE/
    ├── train.json
    ├── validation.json
    ├── test.json
    ├── test_translated_en.json
    ├── test_translated_zh.json
    ├── test_translated_de.json
    ├── test_translated_ar.json
    ├── test_translated_id.json
    └── images/
        └── <image_id>.jpg

Data fields

Field	Type	Description
q_id	integer	Question identifier.
a_id	integer	Answer identifier associated with the question.
question	string	Original veterinary question.
gold_answer	string	Reference answer provided by a verified veterinarian.
gold_question_type	string or null	Question-type label. Manually annotated for PetQA-Bench.
gold_disease	string or null	Comma-separated clinical-condition labels for diagnosis questions.
animal_type	string	Target animal: dog or cat.
image	string	Image identifier.
question_<language>	string	Translated PetQA-Bench question.
gold_answer_<language>	string	Translated PetQA-Bench reference answer.

The translation suffix is one of en, de, zh, ar, or id.

Examples

Text-only

{
  "q_id": 1,
  "a_id": 0,
  "question": "...",
  "gold_answer": "...",
  "gold_question_type": "진단",
  "gold_disease": "...",
  "animal_type": "dog"
}

Multimodal

{
  "q_id": 2,
  "a_id": 0,
  "image": "<image-id>",
  "question": "...",
  "gold_answer": "...",
  "gold_question_type": "진단",
  "gold_disease": "...",
  "animal_type": "cat"
}

Dataset usage

from datasets import load_dataset

data = load_dataset('humane-lab/PetQA')

The dataset is shared under the CC BY-NC-ND 4.0 license.

When using PetQA, please cite:

@misc{kim2026petqa,
  title         = {PetQA: Benchmarking Veterinary Knowledge and Clinical Reasoning},
  author        = {Taegyun Kim and Youngwook Ham and Jungwook Rhim and
                   Ju-Hyun An and Sungkyu Park and Kunwoo Park},
  year          = {2026},
  eprint        = {2609.04598},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.04598}
}