For those who haven't read it, Rich Caruana's thesis on multi-task learning is beautifully written (the cited 1998 paper here). It's amazing to see how far the field has come, and, at the same time, how advanced the thinking was in the 90s too.
Any reason you decided to go with a token vocabulary size of 256k? Smaller vocab/vector sizes like most models in this size seem to be using (~16-32k) are much easier to work with. Would love to understand the technical reasoning here that isn't detailed in the report unfortunately :(.
The "RAG" part (over the training set) is by far the smallest contribution to the performance gains reported (see the ablation study in section 5.2). I don't think the model is actually learning in-context from the selected samples, but rather is continuing to be better conditioned to sample from the right part of the pre-training distribution here, which does a slightly better job when the samples are on topic (vs. random)