MiquMaid v2-alpha is trained on 1 epoch for 18h running on 2xA100 80GB on the following datasets:

Aesir 1, 2, 3
LimaRP
Norobots
ToxicQA
ToxicGPT (v1, for DPO, we needed the same data in the training and the DPO pairs)

The training isn't finished yet, that's why there is only one GGUF file available.

We will probably wait for a second epoch, then apply a DPO train for uncensoring, and finally a DPO train for RP reply (maybe?) if possible.

The prompting format changed too, we used NsChatml who only use special token to use the role of SYSTEM, USER and BOT.

The EOS stay the same than Chatml vanilla.

Here is the files to import on SillyTavern to use this model with the right prompting:

You can replace the system prompt by your favorite one, it's just an exemple.

Prompt format: NsChatml

<|im_system|>
{sysprompt}<|im_end|>
<|im_user|>
{input}<|im_end|>
<|im_bot|>
{output}<|im_end|>