2.13.4 (2026-05-16)

Documentation

• docs: Update documentation for caching by default in task.calculate_descriptive_statistics() (#4676)

• Update documentation for caching by default in task.calculate_descriptive_statistics()

• lintter

• remove extra logging

• fix indentation

• Apply suggestions from code review

Co-authored-by: Kenneth Enevoldsen <kenevoldsen@pm.me>

Co-authored-by: Kenneth Enevoldsen <kenevoldsen@pm.me> (0884d1e)

Fix

• fix: add AMI as clustering metric (#4654)

• feat: always compute AMI alongside V-measure in clustering

Adds Adjusted Mutual Information (AMI) as a clustering metric alongside V-measure. Both metrics are computed on every bootstrap iteration via the _METRIC_FUNCS registry, so tasks get both scores without any opt-in.

The v_measures result key is retained for backward compatibility; AMI scores are exposed under ami_scores / ami / ami_std.

Motivated by review on PR [#4609] (HumanConceptsClustering), whose source paper (arXiv:2505.17117) uses AMI as its main metric.

• refactor: move metric_funcs inside _evaluate_clustering_bootstrapped

The mapping is only used inside the bootstrap loop, so scoping it to the function reads more naturally and keeps the module top-level lean.

• test: cover v_measure + ami in clustering evaluation

Adds focused tests for the clustering evaluation path:

• Direct unit tests on _evaluate_clustering_bootstrapped: registry returns both metrics, perfect labels score near 1, random labels give AMI near 0 (chance correction).

• End-to-end test that runs mteb.evaluate on MockClusteringFastTask and asserts the result dict exposes v_measure, v_measure_std, v_measures, ami, ami_std, and ami_scores.

• simplify

Co-authored-by: Kenneth Enevoldsen <kennethcenevoldsen@gmail.com> (26a44b3)