All sources were originally written in or translated to Russian with Deepseek-v4-pro/Deepseek-v4-flash. Only ~2% of the dataset contains <think> blocks — so the model ships in no-think mode, and it's recommended to run it without thinking.
Recommended settings
Param
Value
Temperature
0.65
Min P
0.05
Top P
0.95
Repetition penalty
1.05
Format
ChatML
Roleplay style: speech in plain text, actions *in asterisks*.
SillyTavern
Import the ChatML preset (context + instruct + system prompt, no-think baked in) via Advanced Formatting → Master Import:
