BioInteract: Scalable Benchmark for Evaluating Vision-Language Models under Semantic Variation

Description

BioInteract comprising richly annotated images depicting interactions between organsims, or biotic interactions, provides a natural testbed for tasks involving images and unconstrained, free-form natural language, as interacting organisms are discerned from images alone and their relationship can be expressed through multiple linguistic forms.

BioInteract Dataset

BioInteract, the largest publicly available multimodal dataset of biotic interaction, specifically curated for vision and machine learning application in the context of AI-driven ecological research. BioInteract includes 256K images annotated with 15.4K unique biotic interactions knowledge graphs which represent the semantic relationship between entities as triplets—source taxon, interaction type, target taxon—across five kingdoms (Animalia, Plantae, Fungi, Chromista, and incertae sedis) and nine ecologically standardized interaction types. A key contribution of BioInteract is that it can generate semantically controlled linguistic variations directly from the underlying knowledge graph. By leveraging structured triplets, we can systematically construct both meaning-preserving and contradictory query variants, enabling explicit control over semantic similarity. This allows us to disentangle correctness from consistency and to rigorously evaluate model robustness under targeted linguistic transformations.

Directory

main/
├── BioInteract/
│   └── train-00000-of-00001.parquet
├── BioInteract-benchmark/
│   ├── BioInteractCommon.csv
│   └── BioInteract100.csv
├──README.md
└──.gitattributes