File size: 1,606 Bytes
d51484d 759b2ac d51484d 759b2ac 4fa77a9 759b2ac d920445 759b2ac 43f822e 759b2ac 43f822e 759b2ac |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 |
---
license: other
language:
- en
---
## Model details
MythoLogic-Mini-7b can be considered the little brother in my Mytho series of models: [MythoLogic-13b](https://huggingface.co/Gryphe/MythoLogic-13b) and [MythoBoros-13b](https://huggingface.co/Gryphe/MythoBoros-13b)).
Its Llama-2 core is powered by [Nous Hermes-2](https://huggingface.co/NousResearch/Nous-Hermes-llama-2-7b), which is further augmented by [Stable Beluga](https://huggingface.co/stabilityai/StableBeluga-7B) and a carefully distilled [Kimiko LoRa](https://huggingface.co/nRuaif/Kimiko_7B).
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.
![](MythoLogic-Mini-7b.png)
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](https://huggingface.co/TheBloke/MythoLogic-Mini-7b-GGML) - [GPTQ](https://huggingface.co/TheBloke/MythoLogic-Mini-7b-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:
``` |