nothingiisreal/L3.1-8B-Celeste-V1.5

Model Overview

L3.1-8B-Celeste-V1.5 is an 8 billion parameter model built upon Meta's LLaMA 3.1 8B Instruct, fine-tuned by nothingiisreal. It was trained with an 8K context window using a novel blend of datasets, including Reddit Writing Prompts, Kalo's Opus 25K Instruct, and c2 logs. This version prioritizes high coherency and strong adherence to OOC: (Out of Character) instructions, making it particularly steerable for creative applications.

Key Capabilities

• Exceptional OOC: Instruction Following: The model is highly responsive to OOC: prompts, allowing users to steer its behavior and output style effectively, even deep within a conversation.
• High Coherency and Creativity: Demonstrates increased intelligence, reduced formatting issues, and higher creativity compared to previous versions, making it suitable for complex narrative generation.
• Flexible Writing Styles: Excels at style copying from few-shot examples and can adapt to a wide range of writing styles, including SFW and NSFW content, without explicit jailbreaks.
• Optimized for Roleplaying: Designed to be highly engaging in roleplay scenarios, capable of evolving character personas and driving narratives proactively.

Usage Recommendations

• System Message: A specific system message is recommended to maximize performance, focusing on character persona maintenance, narrative driving, and diverse output types.
• Swiping and Editing: Users are encouraged to swipe 2-3 times for varied responses and to edit initial messages or use OOC: prompts to refine the model's output and steer its behavior.
• Context Handling: While trained on 8K context, the model is expected to inherit longer context capabilities from its LLaMA 3.1 base, with testing indicating potential for up to 16K tokens.