Overview

Epiculous/Crimson_Dawn-V0.1 is a language model based on the Mistral-Nemo-Base-2407 architecture, developed by Epiculous. Its core design philosophy is to offer an AI experience that is both engaging for roleplay and effective as a general assistant, moving beyond generic interactions.

Training Methodology

The model underwent a two-phased LoRA training process across two epochs on dual NVIDIA A6000 GPUs. The initial phase focused on roleplay (RP) data, after which the LoRA was applied to the base model. Subsequently, the modified base was trained for two epochs on instruct data, with a new instruct LoRA applied to achieve the final model.

Prompting and Usage

Crimson Dawn is designed to be prompted using the Mistral Instruct template ("<s>[INST] Prompt goes here [/INST]<\s>"). For quantized GGUF versions, ChatML is recommended due to observed performance improvements. Users are advised to set the tokenizer to API (WebUI/koboldcpp) when using SillyTavern.

Key Differentiators

• Dual-purpose design: Optimized for both immersive roleplay and practical assistant functions.
• Specialized training: A two-phaged LoRA training approach specifically targeting RP and instruct capabilities.
• Prompting flexibility: Supports Mistral Instruct template with a recommendation for ChatML with GGUF quantizations to enhance performance.