ArliAI-RPMax-12B-v1.2: Creative Roleplay Model

ArliAI-RPMax-12B-v1.2 is a 12 billion parameter model from the ArliAI RPMax series, built upon the Mistral Nemo 12B Instruct 2407 base. This model is specifically optimized for creative writing and roleplay (RP), focusing on generating diverse and non-repetitive content.

Key Capabilities & Features

• Enhanced Creativity: Trained on a curated and deduplicated dataset of creative writing and RP examples, ensuring variety in characters and situations.
• Non-Repetitive Output: Designed to prevent the model from latching onto specific personalities or repetitive patterns, offering a fresh and dynamic RP experience.
• Large Context Window: Supports a 128K context length, allowing for extended and complex roleplay scenarios.
• Optimized Training: Underwent 1 epoch of training with a low learning rate and gradient accumulation to minimize repetition sickness.
• Mistral Base: Benefits from the uncensored nature of its Mistral base, making it highly adaptable for various creative prompts.

Training Details

The model was trained for approximately 2 days on 2x3090Ti, utilizing a 64-rank 128-alpha LORA configuration, resulting in roughly 2% trainable weights. The v1.2 update specifically removed non-creative/RP examples and improved dataset filtering to enhance relevance and quality.

Use Cases

This model is ideal for applications requiring highly creative, varied, and non-repetitive text generation, particularly in roleplaying, interactive storytelling, and character development.