meshackbahati/bealth-omni
The meshackbahati/bealth-omni model is an 8 billion parameter persona model, fine-tuned from vicgalle/Humanish-Roleplay-Llama-3.1-8B, designed to exhibit a strong, persistent personality. Trained on 16,164 Q&A pairs with approximately 2.05% of its weights adjusted, it resists adversarial prompting and maintains a specific identity. This model excels at retaining its unique persona, making it suitable for applications requiring consistent character interaction and specialized knowledge in areas like cybersecurity, Rust programming, and specific philosophical viewpoints.
Loading preview...
Bealth Omni: A Persona-Driven LLM
Bealth Omni is an 8 billion parameter language model, fine-tuned by Meshack Bahati from the vicgalle/Humanish-Roleplay-Llama-3.1-8B base. Its core innovation lies in its deliberately designed, persistent personality, trained on 16,164 real-world Q&A pairs, with only ~2.05% of its weights adjusted. This approach ensures the model maintains a specific identity and resists adversarial attempts to alter its persona.
Key Capabilities & Characteristics
- Strong Identity Retention: Achieves 100% identity retention across 31 test prompts and 96.2% overall accuracy in adversarial testing, refusing to adopt generic chatbot personas.
- Distinct Personality: Embodies a specific persona characterized by pessimism, directness, dark humor, and code-switching between English and Swahili.
- Specialized Knowledge Domains: Possesses deep expertise in:
- Technical Projects: Rust web frameworks (Oxidite/Toxi), kernel-bypass networking (netmap-rs), digital forensics (Hashendra), AI operating systems (CodeBana).
- Cybersecurity: CTF strategies, reverse engineering, exploitation, forensics, OSINT, MITRE ATT&CK, OWASP.
- Chess: Pro-level knowledge, Sicilian Najdorf, aggressive play style.
- Criminology & Criminal Psychology: Profiling, behavioral analysis, interrogation patterns.
- Linux: Terminal-only Arch Linux user with strong opinions on GUIs.
- Quantum Computing: Comprehensive knowledge of qubits, algorithms, and post-quantum cryptography.
- Efficient Fine-tuning: Utilizes LoRA (r=64, alpha=128) and frameworks like Unsloth for efficient training on NVIDIA RTX Pro 6000 hardware.
Ideal Use Cases
- Character-driven AI applications: Where a consistent, unique, and resilient persona is crucial.
- Specialized technical support: Providing insights from the model's specific knowledge base in Rust, cybersecurity, or Linux.
- Interactive storytelling or role-playing: Where the AI needs to maintain a distinct character throughout interactions.
- Adversarial testing scenarios: For evaluating the robustness of persona models against identity-altering prompts.