We created Approximate Likelihood Matching, a principled (and very effective) method for *cross-tokenizer distillation*!
With ALM, you can create ensembles of models from different families, convert existing subword-level models to byte-level and a bunch more🧵
6 months ago
