Skip to content

Release notes

CARA follows semantic versioning. Upgrading is described in Deployment.

0.1.0 (1 November 2026)

The first public release, archived as 10.5281/zenodo.23866167.

  • Judging: groups of similar images with a label bar, labels per image, splitting a group, a review round sorted by label, and an optional Suggest mode. Image, text and composed (image and text) queries.
  • Pools: given candidates, the nearest neighbours of one or more encoders, each to its own depth, or a random sample.
  • Grouping: agglomerative clustering, k-means and HDBSCAN, with a cap on group size; new methods are one module each.
  • Encoders: DINOv3, CLIP, SigLIP 2 and any Hugging Face image or text–image model; new encoders are one module each.
  • Study workflow: pilot with speed and agreement measures, annotation guidelines with examples, a settings simulator that compares groupings in time and recall, and assignments for several annotators.
  • Results: Cohen's κ and Krippendorff's α, adjudication, and exports of TREC qrels (per annotator and by consensus), training pairs and the action log.
  • Running it: Docker Compose, with an optional GPU override for encoding and a toy dataset. A script downloads The Met from its authors and prepares a study.