After some testing, finetuning and multiple merges of Llama-3 LLM models, here is something a little different.
This model is a MoE of 2x Llama-3 model trained on different RP format.
This repo contains FP16 files of Chatty-2x8B.
The idea
I started with two separate Llama-3-Instruct-8B models, each fine-tuned for specific RP formats.
Here is two simple exemple of how it was trained.
Expert 1: This model is trained to handle RP that requires actions and descriptions between asterisks. For example:
*nods* Yes, I understand.
Expert 2: This model is fine-tuned for plain text RP where characters’ dialogues and actions are described straightforwardly. For example:
Nods. "Yes, I understand."
My initial idea was to make a 11B or bigger Llama-3 model, or just make a 2x8B from existing model, but I got some issues, they were not stable enough, even after DPO and FFT on top my frankenmerge/moe of Llama-3, it was not working well enough to release them.
So I just tried the idea of having 2 different RP format trained on 2 separated Llama-3-Instruct-8B, and it worked pretty well!
The dataset
Based on Lumimaid 8B OAS success I still used the same "balance" between RP and non RP in the dataset, the maximum was 50% non RP data on each side.
RP data was different with some exception, the non RP data was exactly the same, despite that, I can't produce repetition so the double usage of non RP datasets didn't hurt the model in the end.
