crestf411/MN-Slush

MN-Slush: A Two-Stage Fine-Tuned Model for Creative and Roleplaying Tasks

MN-Slush is a 12 billion parameter model developed by crestf411, built upon the mistralai/Mistral-Nemo-Base-2407 architecture. It employs a distinctive two-stage training approach designed to significantly boost its creativity, writing capabilities, and roleplaying performance.

Key Capabilities & Training Insights

• Two-Stage Training: The model undergoes an initial pretraining continuation stage on the base model, merged into the instruction-tuned version, to enhance creative output. This is followed by a second fine-tuning stage to further refine roleplaying abilities and address any potential degradation from the merge.
• LoRA Dropout: Utilizes high LoRA dropout (0.5) during both stages, a technique motivated by recent research to improve model generalization and performance.
• LoRA+ Integration: Incorporates LoRA+ with a high LR Ratio (15) for efficient and effective fine-tuning.
• Context Length: Trained with a context size of 16384 tokens in both stages, supporting moderately long interactions.
• Merge Method: The final model is a merge of the trained components using the TIES method, combining the base and fine-tuned LoRAs.

Good For

• Creative Writing: Excels in generating imaginative and diverse text.
• Roleplaying Scenarios: Specifically fine-tuned to enhance interactive and character-driven conversations, particularly following Silly Tavern presets (Mistral V2 & V3).
• Generative Applications: Suitable for tasks requiring a model with enhanced imaginative and expressive capabilities.