conditional_spectral_entropy

probly.quantification.conditional_spectral_entropy = <flextype.singledispatch.Flexdispatch object>

Compute the mean spectral entropy within groups of embedding samples [WGD+26].

The outer sample holds the groups, such as clarifications of a prompt, and each inner sample the responses within a group. The Von Neumann entropy of each group’s kernel is averaged over the groups.

Parameters:
  • embeddings – Sample of embedding samples.

  • gamma – RBF bandwidth parameter.

  • normalized – Whether embeddings are L2-normalized.

  • eps – Numerical cutoff for trace and eigenvalues.

Returns:

Mean spectral entropy over the remaining batch shape.