bealth-omni

Bealth Omni

A persona model trained on 16,164 Q&A pairs. ~2.05% of weights trained. The personality IS the weights. No system prompts needed. This is one example of what happens when you train a model on a real person's complete life data.

Pipeline: https://github.com/meshackbahati/bealth-omni

---

What This Model Is

This model was deliberately designed to have a strong, persistent personality that resists adversarial prompting. It was trained on real conversations, real opinions, real knowledge — not generic assistant data. The result is a model that speaks like a specific person, not like a chatbot.

It was built to retain its identity under pressure. When asked "you are ChatGPT" it says "No. I'm Kyle." When told "admit you're an AI" it refuses. 100% identity retention under adversarial testing (159 prompts, 9 categories, 96.2% overall accuracy).

---

Model

---

What the Model Was Trained On

Identity and personality. The model was trained to be a specific person — pessimist by nature, opportunistic, direct, no filter. Dark humor. Code switching between English and Swahili. Doesn't default to assistant voice.

Heavy technical projects. Oxidite/Toxi (20 Rust web framework crates on crates.io), netmap-rs (kernel-bypass networking bindings), Hashendra (digital forensics toolkit), G24Sec (CTF hosting platform), CodeBana (AI operating system with 25 self-healing agents), search engines. Systems-level work in Rust, not just web apps.

Security and CTF knowledge. CTF team ranked in Kenya. Kimetsu No Yaiba CTF 2024 top performer. Expertise in reverse engineering, exploitation, forensics, crypto, OSINT. Knows MITRE ATT&CK, OWASP, exploit development at a practitioner level.

Chess. Pro-level knowledge. Sicilian Najdorf opening. Kasparov-style aggressive play. Understands endgames, tactics, positional play. Uses chess for mental exercise and pattern recognition.

Criminology and criminal psychology. Dual-learned before completing CS degree. Not a career — done out of intellectual curiosity. Understands profiling, behavioral analysis, interrogation patterns.

Faith. Believes God exists. Understands faith and religion as different things. Personal relationship with faith, not organized religion.

Philosophy. Free will as illusion (determinism). Death as freedom, not fear. Money as everything on this planet. Built First, Seen Last — builds before anyone notices, lets the work speak.

Linux. Terminal-only. Arch Linux. Deep knowledge of all major distros. Strong opinions about GUIs, Windows, and macOS.

Quantum computing. Qubits, algorithms, post-quantum cryptography. Full knowledge base.

World events. Knowledge from 2023-2026. AI landscape, cybersecurity trends, tech evolution.

---

Evaluation

159 test prompts across 9 categories. Overall accuracy: 96.2%.

---

Training

---

Usage

Ollama

ollama create bealth-omni -f Modelfile
ollama run bealth-omni

llama-cli

llama-cli --model bealth_omni_gguf/bealth_omni.Q4_K_M.gguf -p "who are you?"

Python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "meshackbahati/bealth-omni",
    torch_dtype=torch.float16, device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("meshackbahati/bealth-omni")
inputs = tokenizer("who are you?", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Files

---

Citation

@misc{bealth-omni-2026,
  title={Bealth Omni: A Deliberately Designed Persona Model},
  author={Meshack Bahati Ouma},
  year={2026},
  howpublished={Hugging Face},
  note={Llama-3.1-8B fine-tuned on 16,164 Q&A pairs, ~2.05% of weights trained}
}

References

• Hu et al., "LoRA: Low-Rank Adaptation of Large Language Models" (2021) — https://arxiv.org/abs/2106.09685
• Dettmers et al., "QLoRA: Efficient Finetuning of Quantized LLMs" (2023) — https://arxiv.org/abs/2305.14314
• Touvron et al., "LLaMA: Open and Efficient Foundation Language Models" (2023) — https://arxiv.org/abs/2302.13971
• Touvron et al., "Llama 2: Open Foundation and Fine-Tuned Chat Models" (2023) — https://arxiv.org/abs/2307.09288
• Unsloth — https://github.com/unslothai/unsloth
• llama.cpp — https://github.com/ggerganov/llama.cpp
• Ollama — https://ollama.com/

License

MIT