Honestly wanted to wait to quant these. It felt like they were pumping out new versions of it like clockwork to the point where I was wondering if MarinaraSpaghetti was even sleeping. But since this seems to be a clear-cut winner on the board, here we go.
My main goal is to merge the smartness of the base Instruct Nemo with the better prose from the different roleplaying fine-tunes. This one seems to be the best out of all, so far. All credits and thanks go to Intervitens, Mistralai, Invisietch, and NeverSleep for providing amazing models used in the merge.
Mistral Instruct.
Lower Temperature of 0.35 recommended, although I had luck with Temperatures above one (1.0-1.2) if you crank up the Min P (0.01-0.1). Run with base DRY of 0.8/1.75/2/0 and you're good to go.
You can use my custom context/instruct/parameters presets for the model from here.
This is a merge of pre-trained language models created using
mergekit.
This model was merged using the della_linear merge method using F:\mergekit\mistralaiMistral-Nemo-Base-2407 as a base.
1models:
2 - model: F:\mergekit\invisietch_Atlantis-v0.1-12B
3 parameters:
4 weight: 0.16
5 density: 0.4
6 - model: F:\mergekit\mistralaiMistral-Nemo-Instruct-2407
7 parameters:
8 weight: 0.23
9 density: 0.5
10 - model: F:\mergekit\NeverSleepHistorical_lumi-nemo-e2.0
11 parameters:
12 weight: 0.27
13 density: 0.6
14 - model: F:\mergekit\intervitens_mini-magnum-12b-v1.1
15 parameters:
16 weight: 0.34
17 density: 0.8
18merge_method: della_linear
19base_model: F:\mergekit\mistralaiMistral-Nemo-Base-2407
20parameters:
21 epsilon: 0.05
22 lambda: 1
23 int8_mask: true
24dtype: bfloat16
