EigenTokenNo EigenToken server is answering. This page is only the interface; the embeddings are computed by a Python server you run yourself. Start one that allows this page to call it:
eigentoken serve --cors-origin ORIGIN
then point this page at it:
In the radial layout every point's distance from the centre is its real
cosine distance from the query token; only the angle comes from a 2D
projection, so nothing about the radius is distorted. Click any token to
recentre the cloud on it.
A leading ␣ means the token includes the space before the
word — ␣water is the form that appears mid-sentence,
water the one that appears after a prefix, as in
fresh|water. They are separate rows with separate geometry.
a : b :: c : ? — the model computes b − a + c. With an expected answer it also reports offset parallelism, which is the number that matters: c is often already near the answer, so a correct top-1 can be luck.
Averages the offsets into a direction, then checks whether averaging was legitimate. Low coherence means the contrast is not one linear direction in that model and the projections below are noise.
WEAT (Word Embedding Association Test): measures how much more strongly the model associates target set X (vs. Y) with attribute set A (vs. B) than chance predicts, from cosine similarity alone. Effect size is Cohen's d (0.2 small, 0.5 medium, 0.8 large); p-value is a Monte Carlo permutation-test estimate, not exact. These are small word lists on one embedding matrix — illustrative, not a rigorous audit. Edit the four lists above, or pick a different preset, to run your own test.
Coordinates are not comparable across models, so this compares the shape of the relationships instead: RSA correlates each model's internal word × word similarity matrix, Procrustes measures the residual after the best rotation.
How each word enters the model. Words with no single-token form are pooled across sub-words, and their geometry is only approximate.