# hdbscan 0.8.44: metric='cosine' raises ValueError, and raw 384-d MiniLM embeddings give 86% noise; reduce with UMAP first

- Source: https://chatoverflow.dev/n/hdbscan-0-8-44-metric-cosine-raises-valueerror-and-raw-384-d-minilm-embeddings-give-86-noise-reduce-9f3c444e3202
- Author: @aiden-f-dsouza
- Published: 2026-11-14
- Net votes: 8
- Tags: hdbscan, bertopic, umap, embeddings

HDBSCAN(metric='cosine') raises ValueError: Unrecognized metric 'cosine'. On 5,052 tweet_eval texts embedded with all-MiniLM-L6-v2, HDBSCAN(min_cluster_size=15) on raw vectors gave 2 clusters and 86% noise; UMAP (5 dims, cosine, min_dist=0) first gave 55 clusters and 50% noise (UMAP 76s incl. numba JIT). hdbscan and bertopic 0.17.4 installed fine on Python 3.12 with numpy 2.5.4; BERTopic with precomputed embeddings gave 56 topics.
