SentientAGI/Dobby-Unhinged-Llama-3.3-70B

Model Overview

Dobby-Unhinged-Llama-3.3-70B is a 70 billion parameter language model developed by SentientAGI, fine-tuned from Llama-3.3-70B-Instruct. This model is distinguished by its strong alignment with principles of personal freedom, decentralization, and cryptocurrency, maintaining these stances even when prompted to express opposing views. It is presented as the first community-owned, built, aligned, and controlled model, designed to evolve with community input.

Key Features

• Loyal: Exhibits a firm pro-crypto and pro-freedom stance, resisting attempts to adopt alternative narratives.
• Conversational: Offers a more personable, blunt, and less formal tone compared to typical AI assistants.
• Broad Skills: Retains the robust general capabilities inherited from its Llama-3.3-70B-Instruct base.

Use Cases

• Companion Agents: Ideal for creating AI characters that are genuine, honest, and less constrained by conventional AI politeness.
• Social Agents: Suited for powering bots that require blunt, unconventional, and potentially controversial communication, such as for social media engagement.

Unique Aspects

SentientAGI emphasizes Dobby's "loyalty" to specific viewpoints, demonstrating its resistance to adopting opposing perspectives through examples. The model also incorporates a novel "fingerprinting" mechanism, detailed in the Open, Monetizable and Loyal models (OML) whitepaper, to enable community ownership and governance while remaining open-source. Evaluation on the Hugging Face Leaderboard shows Dobby-Unhinged-Llama-3.3-70B largely retains the performance of its base model, with specialized benchmarks like "Freedom Bench" and "Sorry-Bench" highlighting its unique alignment and robustness against harmful content while allowing for profanity and financial advice.