mteb-classification-logistic-regression-100-iterations
IN premise — summaries/2026/08/24/muennighoff-2022-mteb-s3-the-mteb-benchmark.md
Created 2026-08-25T02:58:19+00:00
MTEB classification evaluation trains a logistic regression (max 100 iterations) on frozen embeddings; the embedding model is not fine-tuned.
Summary
The MTEB classification score measures how well the embedding vectors alone separate different text categories, using a deliberately simple and short logistic regression probe rather than a fully trained model. Because the embeddings are frozen and the classifier is capped at 100 training steps, the number reflects the quality of the representation itself, not how well a downstream model can be optimized.