Expert saturation and "ensuring smartness"
#1
by
karan4d
- opened
Curious about this snippet in the card:
Effort was put into ensuring that the smartness of Nous Hermes 2 remained unaffected while also allowing the user to apply a pantheon of personalities to act as 'skins' of a sort.
Would love some insight into reproducibility here. We've just recently started playing with MergeMonster.
Oh, it's nowhere near as impressive as it sounds! I mainly ensured that the system prompts were simple yet distinct and that the resulting dialogue included character names.
My theory is that with those two steps the entire model is less likely to be influenced by the influx of new training data due to how "specific" the data becomes.