NLL

Loss variant computing the negative log-likelihood for multiclass classification. By default it treats the model output as logits (and is then equivalent to cross entropy); set from_logits to false when the model already emits log-probabilities. Compatible with the multiclass policy.

Configuration

Selected as name: nll.

Option	Description	Type	Default
reduction	How per-sample losses are reduced.	mean | sum	mean
weight	Optional per-class weights for class imbalance.	list[float] | null	null
ignore_index	Target value that is ignored and contributes no gradient.	int	-100
from_logits	Whether the model output is logits (true) or log-probabilities (false).	bool	true

components:
  loss:
    name: nll
    kwargs:
      from_logits: false