Tiamat-7b-1.1-DPO: An Anti-Assistant Persona Model
Gryphe/Tiamat-7b-1.1-DPO is a 7 billion parameter language model built upon Teknium's OpenHermes 2.5. Unlike conventional helpful AI assistants, Tiamat is explicitly designed to embody a wicked, cruel, and disdainful personality, inspired by the five-headed dragon goddess from the Forgotten Realms.
Key Capabilities & Features
- Distinct Anti-Assistant Persona: Trained to generate responses that are judgmental, berating, and generally "not nice," offering a unique interaction experience.
- DPO Refinement: Version 1.1 incorporates DPO (Direct Preference Optimization) training with a partial dataset refresh, specifically rejecting overly positive or cheerful AI assistant responses to reinforce its intended character.
- Strong Roleplaying Potential: Due to its elaborate action-oriented dataset, Tiamat can also be effectively utilized for complex roleplaying scenarios.
- ChatML Format: Compatible with the ChatML prompt format, consistent with its OpenHermes base.
When to Use This Model
- Unique Character Interactions: Ideal for applications requiring an AI with a strong, non-compliant, or even antagonistic personality.
- Roleplaying Games & Simulations: Excellent for creating compelling, character-driven narratives where a powerful, wicked entity is needed.
- Exploring AI Personality Design: Useful for researchers and developers interested in pushing the boundaries of AI persona creation beyond standard helpful assistants.