---

name: umap-learn description: Reduce and visualize high-dimensional representations for CS experiments. Use when exploring embeddings, hidden states, feature vectors, retrieval spaces, clustering structure, error slices, or representation drift in model outputs and benchmark data.

UMAP-Learn

Use this skill when a lower-dimensional view can reveal structure that summary statistics hide.

Good Fits

• visualize embedding spaces from NLP, vision, or multimodal models
• inspect class or error clusters
• compare representations across checkpoints or training stages
• preprocess features before clustering or nearest-neighbor analysis
• explore dataset shift or failure slices

Local References

• references/api_reference.md

Working Rules

• Standardize or normalize inputs when the metric requires it.
• Do not over-interpret 2D layouts as proof of causal structure.
• Compare multiple seeds or parameter settings when the geometry matters.
• Use labels, slices, or metadata overlays to interpret clusters responsibly.

Default Output

When the user asks for UMAP help without more detail, return:

1. recommended preprocessing
2. parameter considerations
3. visualization and interpretation notes
4. likely pitfalls