Concept
Reverse-perplexity curriculum
Reverse-perplexity curriculum is a way of ordering supervised fine-tuning data by how surprised the model is by each target trajectory, then feeding those examples in a deliberate sequence rather than shuffling them. Perplexity here is the exponentiated per-token loss of a…
The rest of “Reverse-perplexity curriculum” is a premium feature: every concept in the library gets a precise, practitioner-focused write-up like this one, cross-linked straight from the paper summaries that use it.
Log in to unlock→