Model Overview

crestf411/L3.1-nemotron-sunfall-v0.7.0 is a specialized language model developed by crestf411, created by merging and training with NVIDIA's Nemotron 70B Instruct. This version, v0.7.0, features a significantly revamped dataset metadata generation process aimed at improving overall data quality.

Key Capabilities

• Interactive Storywriting: The model is specifically trained for interactive story generation, allowing users to steer narratives by providing scenario prompts.
• Roleplay Optimization: It incorporates datasets like Gryphe/Sonnet3.5-Charcard-Roleplay and various Antracite.Org datasets, enhancing its ability to engage in character-driven roleplay.
• Silly Tavern Compatibility: Training included context mimicking Silly Tavern's "Llama 3 Instruct" preset, making it well-suited for such environments.
• Advanced Training Techniques: Utilizes "Fine-Tuning with Very Large Dropout" (LoRA dropout of 0.5) and a constant learning rate of 4e-6, with a modified alpha setting during merging to retain Nemotron's intelligence.

Recommended Usage

Optimal performance is achieved with specific inference parameters:

• Temperature: 1
• MinP: 0.05-0.1
• DRY: 0.8 1.75 2 0

For higher context sizes (10k+), enabling XTC (e.g., at 0.1 / 0.5) can significantly improve output quality, especially if the model's responses become uninspiring. Conversely, if the model makes subtle or obvious factual mistakes, lowering the temperature or disabling XTC is recommended.