Evaluating diffusion language models can be tricky… alongside reporting gen. PPL and entropy we suggest you try SOL 🌞
Recently there has been a lot of discussion about the difficulties of evaluating diffusion language models.
We instead introduce: SOL, a sample based distance between generated and real text distributions. 🧵View quoted post on X
