Not-For-All-Audiences
nsfw
Exl2 version of Undi95/OpenDolphinMaid-4x7b
branch
7bh8 : 7bpw h8
Using ThePile 0007.parquet as dataset
Quantization settings : python convert.py -i models/Undi95_OpenDolphinMaid-4x7b -o OpenDolphinMaid-4x7b-temp -cf OpenDolphinMaid-4x7b-7bpw-h8-exl2 -c 0007.parquet -l 8192 -b 7 -hb 8 -ml 8192
below this line is original readme
Merge of OpenHermes and Dolphin with 2x Noromaid DPO, trying to add a little more brain in the model, while being smaller than a 8x7b.
It seems to work well.
Description
This repo contains fp16 files of OpenDolphinMaid-4x7b.
Models and LoRA used
- NeverSleep/Noromaid-7B-0.4-DPO x 2
- teknium/OpenHermes-2.5-Mistral-7B
- cognitivecomputations/dolphin-2.6-mistral-7b-dpo
Prompt template: Chatml
<|im_start|>system
{sysprompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
If you want to support me, you can here.