Model turns in PJMixers/grimulkan_bluemoon_Karen_cleaned-carded-formatted regenerated using tngtech/DeepSeek-TNG-R1T2-Chimera.
You should mask everything except the last turn when training. All previous model turns are the original dataset.
It's setup to be trained like R1:
NousResearch/Minos-v1 was used to avoid refusals. Only checked against <|user|>\n{latest_user_turn}\n<|assistant|>\n{response_without_thinking}, regenerating if not at least 80% confident it's a non-refusal.
