crestf411/L3.1-70B-sunfall-v0.6.1

Model Overview

crestf411/L3.1-70B-sunfall-v0.6.1 is a 70 billion parameter model based on Meta's Llama-3 70B Instruct, featuring a 32768 token context window. This version is specifically fine-tuned for advanced character roleplay and creative story writing, incorporating unique training methodologies.

Key Capabilities & Training

• Immersive Roleplay: Trained to excel as an "expert actor" fully immersing into character roles, mimicking Silly Tavern's Llama3-instruct preset.
• Creative Storytelling: Optimized for generating compelling stories, utilizing a specialized system message for scenario-based narrative creation.
• Lore Book Integration: Designed to work effectively with lore book tags, supporting complex character and world-building.
• "Diamond Law" Adherence: Training data incorporates adherence to a specific set of rules, referred to as the "Diamond Law," influencing model behavior.

Performance & Recommendations

While the model shows an overall MMLU-Pro benchmark improvement (60.73%) over the Llama3.1 70B Instruct base (58.64%), it exhibits variations across categories, with notable gains in biology, engineering, and history. For optimal performance, specific inference parameters are recommended:

• Temperature: 1.2
• MinP: 0.06
• Optional DRY: 0.8 1.75 2 0

Use Cases

This model is particularly well-suited for:

• Interactive Fiction & Roleplaying: Ideal for applications requiring deep character immersion and dynamic narrative generation.
• Creative Writing Assistance: Can be leveraged for generating story plots, character interactions, and descriptive passages.
• Silly Tavern Environments: Specifically designed and tested for enhanced performance within the Silly Tavern platform.