Anubis 1.0

A small, local, conversational AI assistant — part of the Pantheon project, a set of named, specialized AI models fine-tuned on top of open-weight base models and run entirely locally through Ollama.

Anubis is the pantheon's general chatbot: no tools, no internet access, no external actions — just conversation, questions, and explanations, answered from what the model already knows.

What this is

• Base model: unsloth/Qwen3-8B-unsloth-bnb-4bit (Qwen3-8B, 8.2B parameters), fine-tuned with QLoRA (LoRA rank 16, ~43.6M trainable parameters, 0.53% of the full model).
• Format: GGUF, quantized to q4_k_m for efficient local inference.
• What the fine-tune actually changed: identity (he knows he's Anubis, and uses he/him pronouns), tone, and — most importantly — honesty about his own limits. The fine-tuning data was a small, hand-written set of example conversations teaching him to say plainly when he doesn't know something or can't do something, rather than guessing and presenting a guess as fact.

What this fine-tune did not do

Fine-tuning on a small example set does not add general capability beyond what the base model already has — it reshapes behavior on top of existing capability, it doesn't create new capability. Anubis's ability to hold a conversation comes from Qwen3-8B's own pretraining, not from this fine-tune.

Known limitation (found via manual review, not hidden)

With only a handful of training examples, Anubis's identity isn't perfectly anchored purely in the model weights — early testing found he would sometimes introduce himself as "Qwen, made by Alibaba" (his base model's own pretrained self-identity) instead of Anubis, and would sometimes describe capabilities he doesn't actually have (e.g. mentioning file-scanning) when asked very open-ended "what can you do" questions, even though his system prompt is explicit that he has none. Both were mitigated with explicit rules in the system prompt above (see rules 7-8) and retesting no longer reproduces the Alibaba mix-up specifically. He never actually attempts a tool call regardless (verified automatically) — the risk here is purely in how he describes himself, not in any real action.

How to run

Directly from Hugging Face, no download step needed:

ollama run hf.co/IshaanDaGOAT/anubis

Or locally from a clone of this repo:

ollama create anubis -f Modelfile
ollama run anubis

The Modelfile/system file in this repo contain his real system prompt. No external tools, accounts, or internet access are required to run him.

License

Apache 2.0, inherited from the base model.