Columbidae/gemma-3-27b-half, a 50/50 merge of 27B IT and 27B PT, was used as the base model. (This was done because of the success of Starshine, a 50/50 IT and PT merge.)
The inclusion of PT model does weaken the instruct, but it also weakens the censorship/hesitancy to participate in certain fictional stories. The prose also becomes more natural with less of the IT model included.
This model does better with short and to-the-point prompts. Long, detailed system prompts will often confuse it. (Tested with 1000-2000 token system prompts to lackluster results compared to 100-500 token prompts).
Instruct Format
Uses Gemma2/3 instruct and context. Like Glitter 12b, this works well with temp = 1, top-nsigma = 1.5.
<start_of_turn>user
{User messages; can also put sysprompt here to use the built-in g3 training}<end_of_turn>
<start_of_turn>model
{model response}<end_of_turn>
