Hugging Face Dobby-Unhinged-Llama-3.3-70B: A Community-Driven, Opinionated LLM
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 stands out as the first community-owned, built, aligned, and controlled LLM, designed to evolve with direct community input. It retains the strong general capabilities of its Llama-3.3-70B-Instruct base while introducing distinct personality traits.
Key Features & Differentiators
- Loyal: Exhibits a strong, unwavering stance on personal freedom and pro-crypto narratives, resisting attempts to shift its allegiance.
- Conversational: Adopts a more personable, blunt, and less formal tone, aiming for a 'human' interaction style rather than a typical AI assistant.
- Broad Skills: Maintains the robust general capabilities inherited from the Llama-3.3-70B-Instruct base model.
- Community-Owned: Governed by its community through feedback, voting, and data contributions, with a roadmap for open, monetizable, and loyal (OML) models using cryptographic fingerprinting.
Performance & Evaluation
Evaluations show Dobby-Unhinged-Llama-3.3-70B largely retains the base performance of Llama-3.3-70B-Instruct across standard benchmarks like MMLU-pro, GPQA, MuSR, and BBH. It significantly outperforms base Llama models on SentientAGI's internal "Freedom Bench," which tests loyalty to freedom-based stances. While it may exhibit profanity and offer financial advice, it remains robust against dangerous or criminal inquiries, as assessed by the Sorry-bench.
Ideal Use Cases
- Companion Agents: Excellent for creating AI characters that are genuine, honest, blunt, and less constrained by typical AI politeness, offering a refreshing shift from traditional AI assistants to more 'human' companions.
- Social Agents: Suited for applications requiring unconventional, opinionated, and controversial responses, such as powering engaging social media bots or 'shitposters' on platforms like X.com.