This model uses a qlora applied to Mixtral Instruct made with a dataset similar to Libra-19B (totaling roughly 65 megabytes) consisting of raw text samples up to 512 tokens in length, geared towards giving the model a much more visceral output when engaging in role play.
It was trained via qlora using 2xA100 on runpod.
EDIT:
It was originally stated that it also used the synthetic samples used to train Libra-32B but upon reviewing the dataset used this is not the case.
The training was done at a very low learning rate for 3 epochs, aiming for more of a stylistic nudge that leaves the general functionality of the model intact.
Overall while the model is a little more unruly than a pristine copy of Mixtral Instruct it feels overall better at characterization and visual descriptions.
It seems to do well with the Liminal Drift preset on SillyTavern.
While it has less refusals than Mixtral Instruct some of its refusals can be overcome simply with a regeneration.
The model still uses the Mixtral Instruct prompt format e.g.
