Comment by upghost

Yeah, this checks out. I wonder what they are doing to prevent semantic collapse. Also, I wonder if the base model would already be instruct and RLHF tuned or only pre-trained. Trying to do additional training without semantic collapse in a way that is meaningful would be interesting to understand. Presumably they are using adapters but I've never had much luck in stacking adapters.

i.e.:

1. Do I start with an RLHF tuned model, "pretrain" on top of that (with adapter or by freezing weights?), then SFT on top of that (stack another adapter, or add layer(s) and freeze weights?) (and where did I get the dataset? synthetic extraction from corpus?), then RL (adapter, add layer(s) and freeze?)

2. or do I start at SF tuned model, ...

3. or do I start at raw pre-trained model, ...

Would love to know what the matrix used was.