- OpenMidnight foundation model
- User-Agent header to download requests
- H0-mini foundation model
- Minor change to ranking in leaderboards (same rounded performance = same rank)
- Updated timm version requirements to adjust to keep
- HistAI SPIDER datasets (breast, colorecal, skin, thorax)
- Zero-shot VLM classification task
- DINOv3, GIGAPATH, KAIKO-ViT foundation models
- Option to benchmark a model for a task on a custom dataset
- RandStain transformation for the transformation invariance task
- SPIDER and zero-shot leaderboards + up-to-date rank-sum leaderboard
- Option to divide bracs images into patches when extracting embeddings
- Use-case examples in documentation
- Saving a config file at the end of a benchmark run
- Extended support to python versions >=3.10,<3.14
- Updated dependency requirements and replaced setup.py with pyproject.toml
- Loading all embeddings and labels as numpy arrays in dataset init for linear_probing+embedding_pre_loading
- Config setting for probe training with custom model
- Hyper-parameters printing (colors and model name when custom)
- Links to leaderboard, docs, arXiv paper
- Download of esca, tcga_uniform and wilds datasets
- Handling of ignored pixels in dice loss
- First official release
- Preprint