Sao10K/MN-BackyardAI-Party-12B-v1

Overview

Sao10K/MN-BackyardAI-Party-12B-v1 is a 12 billion parameter model, a variant of the private Lyra-v4 architecture (specifically 12B-Lyra-v4a2), fine-tuned for group-chat based roleplaying. It was trained on a unique, entirely human-based dataset derived from forum and internet group roleplaying styles, with LLM augmentation only for character sheet formatting.

Key Capabilities

• Multi-Character Roleplay: Designed to handle group chat scenarios with 2 to 7 unique characters per entry, simulating real-world group roleplay.
• Human-Centric Data: Trained on authentic human roleplay conversations, preserving the quality and uniqueness of human interaction.
• Flexible Turn Structure: Supports disjointed turns, allowing for multiple model replies or user inputs in sequence, mimicking dynamic chat environments.
• Character Card Integration: Optimized for detailed character sheets, with a recommended format to ensure consistent character portrayal.

Use Cases

• Group Roleplaying: Ideal for applications requiring complex, multi-character interactive narratives.
• Dynamic Chat Simulations: Suitable for scenarios where conversational flow doesn't strictly adhere to user -> model -> user turns.
• Interactive Storytelling: Can be leveraged for generating rich, character-driven stories in a group context.

Important Considerations

• Formatting: Utilizes a variant of ChatML with [INST] blocks for multi-character roleplay and standard ChatML for 1-on-1 interactions. Users should set both <|im_end|> and [INST] as stopping strings.
• Limitations: Known issues include occasional impersonation, varied output quality due to diverse human data, potential character detail confusion with too many characters or complex prompts, and a fantasy bias in its dataset. The dataset currently comprises 4,000 varied samples.