How would we go about increasing the fidelity
#8
by
ericleigh007
- opened
As it stands, this model is great for getting ideas that would still need to be developed, because of the output lower fidelity.
I assume the problem here is that the token Encoder that is used during training was decidely low-fi, and probably a desire to keep the context window smaller, and toward training and inference without a massive farm.
Am I correct? Do we need to completely re-train with better quality samples, or is the problem something else?
Is there a happy medium we could reach where fidelity is much higher without blowing out the resource load?
Let's say that we want a local model that's high fidelity, that we can execute on one or two DIGITS computers, for instance.... <3