embeddings map — atlas

A projection of atlas — read it with the provenance below, not without it.

Provenance

Atlas
atlas
Model
Qwen/Qwen3-Embedding-0.6B
Endpoint
http://localhost:8001/v1
Items
152 items in 1024 dimensions
Algorithm
PCA (thin SVD of the mean-centred matrix)
PC1 explained variance
10.5%
PC2 explained variance
9.0%
Total kept
19.5% of the variance in this atlas
Seed
0
Colour groups (k)
5
Stability runs
5 seed rerun(s), 20 resample(s)
Sign convention
largest-magnitude loading forced positive
Aspect
uniform scale on both axes

Colour

Colour is a k-means partition into 5 groups at seed 0, stability verdict: unstable (adjusted_rand_index, 5 seed rerun(s), 20 bootstrap resample(s)). k-means always returns exactly the k it was given, structure or no structure, so the verdict — not the picture — is what says whether these groups are real.

  • cluster 0 · 18 items
  • cluster 1 · 40 items
  • cluster 2 · 23 items
  • cluster 3 · 37 items
  • cluster 4 · 34 items

What this picture does not show

This is a lossy 2-D projection of a high-dimensional embedding space. Only the share of variance reported above survives the projection; the rest was discarded to fit the page. Distances you can see here are not faithful to distances in the space -- especially the distance between apparent clusters, which a projection is free to rearrange. A gap on this page is not evidence of a gap in the model. The layout is conditioned on the algorithm, its parameters, the item set, and (for the colouring) the seed: change any of them and the picture changes, so a structure visible only at these settings is a finding about the settings, not about the model. Read this picture only together with the provenance above; separated from it, a screenshot asserts more than the measurement supports. See docs/design-notes/semantic-measurement.md section 3.