Gryphe/Pantheon-RP-1.5-12b-Nemo: Enhanced Roleplay and Persona Generation
Gryphe/Pantheon-RP-1.5-12b-Nemo is the latest iteration in the Pantheon model series, a 12 billion parameter model built on the Mistral Nemo base. It is specifically designed to introduce a wide collection of distinct personas, enhancing general roleplay experiences through simple activation phrases.
Key Capabilities
- Multi-stage Fine-tuning: The model underwent a two-stage fine-tuning process. The first stage balanced instruct and roleplay data using a subset of the Deduped Sonnet 3.5 SlimOrca dataset and various roleplay sources. The second stage integrated a rebuilt and expanded Pantheon Roleplay dataset, also balanced with a 50/50 ratio, alongside a companion roleplay dataset to ensure broad non-Pantheon roleplay capabilities.
- Diverse Personas: Features a significantly expanded and improved collection of personas, each with unique introductions and system prompts. Examples include Clover (Southern centaur), Raza (raptor scientist), Stella Sabre (batpony Lunar Guard), and returning characters like Aiva (AI companion), Haru (harpy), Kyra (wolfgirl), Nyaa (tabaxi), Nyx (dragon girl), Sera (serpent girl), Tiamat (dragon goddess), Tsune (kitsune), and Xala (shapeshifter).
- Optimized for Roleplay: The second fine-tuning stage specifically focused on an asterisk-style, no-quotes for speech roleplay style (Markdown), leading to a bias towards this format in responses. It supports flexible character card formatting.
- Aiva the Assistant: Includes a distinct assistant persona, Aiva, trained on a mix of instructional and roleplay data, capable of tasks like DM world-building, Python coding, and RSS summarization.
Good For
- Interactive Storytelling: Ideal for applications requiring rich, character-driven narratives and dynamic interactions with diverse personas.
- Roleplay Scenarios: Excels in generating responses in a specific asterisk-style roleplay format, catering to users who prefer this style.
- Character Development: Useful for developers looking to integrate models with pre-defined, complex character archetypes that can be easily summoned and engaged.