Skip to content

ccaligned

ccaligned

CCAligned perplexity evaluation.

Measures language model perplexity on multilingual parallel text. Returns 1/perplexity (higher is better).

Provides per-language eval classes so each can be tracked independently across continual learning phases. All ~90 CCAligned language pairs are registered as ccaligned_<lang> (e.g. ccaligned_fr_xx).

CCAlignedEval(num_samples: int = 500)

Bases: PerplexityEvaluation

Perplexity evaluation on CCAligned multilingual sentence pairs.