|
--- |
|
license: cc-by-nc-4.0 |
|
tags: |
|
- merge |
|
--- |
|
### Chatty-2x8B |
|
|
|
## Description |
|
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 GGUF 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. |
|
|
|
## Prompt template: Llama3 |
|
|
|
``` |
|
<|begin_of_text|><|start_header_id|>system<|end_header_id|> |
|
|
|
{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|> |
|
|
|
{input}<|eot_id|><|start_header_id|>assistant<|end_header_id|> |
|
|
|
{output}<|eot_id|> |
|
``` |
|
|
|
## Others |
|
|
|
Undi: If you want to support us, you can [here](https://ko-fi.com/undiai). |
|
|
|
IkariDev: Visit my [retro/neocities style website](https://ikaridevgit.github.io/) please kek |
