nothingiisreal/L3-8B-Celeste-V1.2
nothingiisreal/L3-8B-Celeste-V1.2 is an 8 billion parameter LLaMA 3 Instruct-based model, fine-tuned by nothingiisreal with an 8K context length. It is specifically optimized for multi-turn and longer context roleplaying, demonstrating increased coherency and style copying abilities. The model excels at creative roleplaying and character interaction, trained on a diverse dataset including Reddit Writing Prompts and Opus Instruct.
Loading preview...
nothingiisreal/L3-8B-Celeste-V1.2 Overview
L3-8B-Celeste-V1.2 is an 8 billion parameter model based on LLaMA 3 8B Instruct, developed by nothingiisreal. It features an 8K context length and has been fine-tuned using a specialized dataset including Reddit Writing Prompts (both SFW and NSFW), Opus Instruct 15K, and cleaned c2 logs. A key focus during training was to enhance coherency in multi-turn conversations and longer contexts by utilizing longer sequences of data.
Key Capabilities & Differentiators
- Specialized for Roleplay: This model is explicitly designed and optimized for roleplaying scenarios, with instruction-following capabilities outside of roleplay being coincidental.
- Enhanced Coherency: Training on longer data sequences from writing prompts significantly improved its ability to maintain narrative flow over extended interactions.
- "OOC:" Steering: The model is highly responsive to out-of-character (OOC) instructions, allowing users to steer character behavior effectively.
- Creative & Dynamic Roleplay: It excels at creative roleplaying, adapting personas based on recent messages rather than just static system prompts or character cards, and demonstrates strong style copying from few-shot examples.
- Uncensored for RP: The model is uncensored for roleplay, offering flexibility for various narrative styles, including NSFW content when desired, while also being capable of SFW interactions.
Usage Recommendations
- Sampling Settings: Recommended settings include Temperature 1.25, Min P 0.1, and Repetition Penalty 1.1 for optimal performance.
- System Message: While a system message is not strictly required, a detailed one is provided for users seeking specific character guidelines.
- Refusal Handling: The model is trained to sometimes refuse romantic advances for realism, but this can be overridden by editing character messages.
- Context Extension: While trained for 8K context, it can be RoPE'd to 16K context if needed, though 8K is generally recommended for stability.
Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.