Perplexity

Perplexity lives in artifex.generative_models.core.evaluation.metrics.text.perplexity.

Current Contract

• accepts a caller-supplied callable model when you want the metric to compute log probabilities from token inputs
• also supports direct log_probs input at compute time when the caller already owns the language-model forward pass
• lower perplexity remains better

Example

metric = Perplexity(model=language_model, batch_size=32, rngs=nnx.Rngs(0))
result = metric.compute(inputs=token_ids)