incerto.llm.TokenEntropy

incerto.llm.TokenEntropy#

class incerto.llm.TokenEntropy[source]#

Bases: object

Compute predictive entropy at each token position.

Higher entropy indicates higher uncertainty in the model’s prediction for that token. This is the most common token-level uncertainty measure.

__init__()#

Methods

__init__()

compute(logits[, dim])

Compute entropy over token distribution.

static compute(logits, dim=-1)[source]#

Compute entropy over token distribution.

Parameters:
  • logits (Tensor) – Token logits of shape (…, vocab_size)

  • dim (int) – Dimension to compute entropy over (default: -1)

Return type:

Tensor

Returns:

Entropy values of shape (…,)