Pathology VLM Benchmark · MahmoodLab
viper
A benchmark for assessing vision-language models on rat histopathology in veterinary and toxicologic pathology.
What viper is
VIPER is a specialised pathology benchmark curated by experts to evaluate the capabilities of vision-language models. Developed by MahmoodLab, it addresses the need for systematic model evaluation within veterinary and toxicologic pathology, focusing on rodent tissue specimens.
The dataset provides visual question-answering and multiple-choice challenges grounded in haematoxylin and eosin (H&E)-stained histology. It allows computational researchers to gauge how well multimodal models interpret microscopic features and answer diagnostic questions in preclinical pathology settings.
How it was made
The benchmark comprises 1,251 expert-generated questions distributed across 419 H&E-stained rat histology images. It pairs textual questions and multiple-choice choices with corresponding histological imagery.
Data records are provided in Parquet format, supporting multimodal image and text processing pipelines in standard scientific data libraries.
Key facts
- Image count
- 419 images
- Question count
- 1,251 questions
- Organism
- Rat
- Staining method
- Haematoxylin and eosin (H&E)
- Task types
- Visual question answering, multiple-choice
- Primary fields
- Veterinary pathology, toxicologic pathology
- File format
- Parquet
What you can use it for
- Vision-language model evaluation
- Benchmarking multimodal foundational models on their capacity to process, interpret, and explain microscopic findings in animal histology.
- Toxicologic pathology testing
- Measuring model accuracy when answering structured questions regarding rat tissues relevant to preclinical toxicological assessments.
- Visual question answering research
- Evaluating model performance on domain-specific multiple-choice visual question-answering tasks in digital pathology.
Explore the data
This dataset is gated
Its authors require their terms to be accepted before the data can be read. Accept them on the original dataset page. The platform does not accept terms on your behalf, so the rows are not shown here.Strengths and limits
Good at
- All 1,251 questions are curated directly by domain experts.
- Addresses an underrepresented niche in machine learning benchmarks by targeting veterinary and toxicologic histology.
- Integrates multimodal image and text questions formatted for standard data science workflows.
Watch out for
- The image set is limited to 419 rat tissue samples.
- Restricted exclusively to haematoxylin and eosin (H&E) staining, omitting special stains or immunohistochemistry.
- Restricted to rat biology, excluding other animal models or human pathology.
Questions
Can I look inside viper without downloading it?
Not here: its authors gate it behind terms each user must accept on the original page. Its description and statistics are still shown.
What licence is viper under?
Its authors declare other. Read the terms on the original page before reusing the data.
What organisms and stains are represented in the dataset?
The dataset contains H&E-stained tissue sections obtained from rats.
What tasks can be evaluated using VIPER?
It supports multiple-choice visual question answering for multimodal vision-language models.
How many images and questions are included?
There are 1,251 questions across 419 images.
Who generated the annotations for VIPER?
The benchmark questions and annotations were created by pathology experts.
How to cite it
arXiv:2608.26382
Written from viper's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.