OddTheGreat/Cogwheel_24B_V.1

merge

This is a merge of pre-trained language models

Goal of this merge was to improve Core's abilities to russian language and to make it better on 'logical' field.

Model is still great as narrator or setting, it seems like now model notices and use even smallest details in description. Model follows instructions and rules well, it is creative when it needs to be, and "smart" enough. Model sometimes will reply for you, especially if user is mentioned in char card, however it fixes by turning on instruct template or by a few swipes. I tested some my overengineered RU cards, RU really improved, and works good if used as assistant, but still worse than EN for roleplay. ERP was tested too, no problems spotted.

While i tested not fully translated to RU char cards, (model works good with them, but prone to language switch in beginning), i was visited by mad idea: test fully ru card with EN first message. And at least on this model it worked, and worked good, model understands card good enough, and with EN output it give life to my lazy old RU cards.

Tested on Q4_K_M, ~600 replies, T 1.04, xtc 0.1 0.2, Mistral template.