Since 7B models tend to be less capable all-rounders, more emphasis was put on improving the roleplaying aspects for this gradient merge, of which various gradients were benchmarked before settling on the configuration shown below.
In technical terms, the Hermes-2 core starts at 90% strength before fading away completely at the 12th layer level, where Stable Beluga (and Kimiko) handle the more intricate linguistic aspects.
Quantized models are available from TheBloke: GGML - GPTQ (You're the best!)
Prompt Format
Due to its Hermes-2 core this model works best with Alpaca formatting, so for optimal model performance, use:
<System prompt/Character Card>
### Instruction:
Your instruction or question here.
For roleplay purposes, I suggest the following - Write <CHAR NAME>'s next reply in a chat between <YOUR NAME> and <CHAR NAME>. Write a single reply only.
### Response:
