Model Overview

Sao10K/L3-8B-Chara-v1-Alpha is an 8 billion parameter experimental model derived from Llama-3-Instruct, designed specifically for multi-character, group-based roleplaying. This model is a proof-of-concept, notable for being trained entirely on human-generated roleplaying data, including forum scrapes, to avoid any LLM-generated content.

Key Characteristics & Training

• Architecture: Built on Llama-3-Instruct, utilizing its format.
• Data Source: Exclusively human-sourced roleplaying data, ensuring no AI model contamination.
• Training Focus: Optimized for group chat scenarios, where one designated character acts as the human turn, and multiple other characters (2-5 unique per entry) are assigned as GPT-turns to simulate real group interactions.
• Dataset Size: Trained on a relatively small dataset of approximately 3,000 sample entries.

Intended Use & Performance

• Primary Use Case: Best suited for group chat roleplaying, especially when character cards are properly defined for a multi-participant setting.
• Experimental Nature: This is an alpha model and is not expected to outperform more mature, general-purpose models in broad benchmarks.
• Limitations:
  • Performance in one-on-one roleplaying might be affected due to its group chat focus.
  • The use of character names that are multiple tokens instead of single tokens may impact output quality.
  • The dataset, while filtered, may still contain occasional low-quality entries as it's a proof-of-concept without extensive manual curation.

Configuration Notes

For users of SillyTavern, it's recommended to remove square brackets from [{{name}}], [{{char}}], or [{{user}}] within the instruction template to match the model's training format.