Skip to content

Commit eac6756

Browse files
realmarcinclaude
andauthored
Default the embedding projection to PaCMAP (PCA-init, seed 42); keep UMAP as an option (#188)
Per the deep-research recommendation: PaCMAP preserves both local and global structure better and more robustly than UMAP at these dataset sizes. Rows are L2-normalized to mirror the previous cosine metric, PCA-initialized, and the random_state is fixed for reproducibility. UMAP stays available via --method umap. Regenerated the community UMAP with the PaCMAP default (292 communities). Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
1 parent a192416 commit eac6756

6 files changed

Lines changed: 1035 additions & 549 deletions

File tree

0 commit comments

Comments
 (0)