Hugging Face Pantheon-RP-1.8-24b-Small-3.1 Overview
Gryphe's Pantheon-RP-1.8-24b-Small-3.1 is a text-only model specifically designed to elevate roleplaying interactions. It introduces a unique "Pantheon" series of personas, each summonable with a simple activation phrase, enriching roleplay with distinct personalities, accents, and mannerisms that typical language models might struggle to convey. The model was developed using the Small 3.1 Instruct version, chosen for its robust handling of varied user inputs.
Key Capabilities & Training
This iteration of Pantheon combines diverse data sources, including synthetic data from Sonnet 3.5 + 3.7, ChatGPT 4o, and Deepseek. This data underwent an extensive rewriting process to eliminate common AI clichés, aiming for a fresh and immersive experience. The model is trained on three distinct categories of roleplay: Pantheon personas, general character cards, and text adventure scenarios (borrowing from AI Dungeon's Wayfarer project). It primarily uses a second-person perspective, referring to the user as "you," and supports both Markdown and novel-style roleplaying with a 30/70 ratio.
Pantheon Personas
The model features a range of predefined personas like Lyra (sassy dragon girl), Clover (hospitable centaur), Haru (language-challenged harpy), and Kyra (tsundere wolfgirl), among others. These personas are designed to be activated via system prompts, with expanded prompt templates available for optimal experience. Users are encouraged to provide details about themselves and the location to give the persona more context, as the model will generate information if not explicitly provided.
Recommended Use Cases
- Immersive Roleplaying: Ideal for generating dynamic and character-rich roleplay scenarios.
- Character Card Interactions: Excels at maintaining consistent character traits and dialogue based on provided character cards.
- Text Adventures: Suitable for creating interactive narrative experiences, drawing from its text adventure training.
Inference Settings
Recommended inference settings include a temperature of 0.8, repetition_penalty of 1.05, and min_p of 0.05. Using character names in front of messages is suggested to help the model focus on specific characters. The model was trained using the ChatML prompt format.