North-ML1/Aurora-One

Aurora One GGUF

Aurora One is a small from-scratch decoder-only language model. This repository contains GGUF exports for local inference.

This is a custom Aurora architecture exported through a Qwen3-compatible GGUF path. It is not a Qwen model.

Files

• aurora-one-generalization-repair-v4-f16.gguf - recommended GGUF for llama.cpp / LM Studio server API.
• aurora-one-generalization-repair-v4-lmstudio-f16.gguf - alternate export with conditional ChatML template metadata.
• SYSTEM_PROMPT.txt - recommended system prompt.
• aurora_lmstudio_adapter.py - optional OpenAI-compatible middleware for deterministic arithmetic/sorting/live-data fallback/search.

Recommended Prompt Format

Use ChatML:

<|im_start|>system
You are Aurora One. Follow the user's instruction exactly. Be concise by default. Do not invent live facts or pretend to use tools. Only use a database, search, internet, or external tool if the system prompt explicitly says it is available. If the answer is not in your training data and no such access is explicitly available, say exactly: According to my training data, I cannot answer this question reliably. For code-only requests, output only working code.<|im_end|>
<|im_start|>user
Hello!<|im_end|>
<|im_start|>assistant

Recommended stop strings:

<|im_end|>
<eos>
<|end|>

LM Studio

The LM Studio lms chat wrapper can route custom qwen3-shaped GGUFs poorly. Use the LM Studio local server API instead.

lms server start
lms load aurora-one-generalization-repair-v4-f16.gguf --identifier aurora-one --gpu max -c 2048 -y

Call:

http://127.0.0.1:1234/v1/chat/completions

Use model: "aurora-one" and include the system prompt from SYSTEM_PROMPT.txt.

Optional Adapter

For a more useful server deployment, run the included adapter in front of LM Studio:

python3 aurora_lmstudio_adapter.py --listen-port 8088 --enable-search

Then call:

http://127.0.0.1:8088/v1/chat/completions

The adapter:

• handles simple arithmetic deterministically,
• sorts comma-separated numbers/words,
• handles a few common deterministic translation/instruction cases,
• returns the safe fallback for current/live facts unless search is explicitly enabled in the system prompt,
• can use CoinGecko for BTC, wttr.in for weather, and modal.com/pricing for Modal GPU pricing.

For search/live access, include a system prompt sentence such as:

Search/internet/database access is available for current facts.

Known Limitations

Aurora One is a small experimental model. It is not a reliable general assistant by itself. It can fail on arithmetic, exact instruction following, factual recall, translation, and reasoning. For production use, keep deterministic tools/middleware around it.

Publish From Local Folder

From this folder:

hf auth login
hf repo create YOUR_USERNAME/aurora-one-gguf --type model
hf upload YOUR_USERNAME/aurora-one-gguf . .

Or with git-lfs:

git init
git lfs install
git remote add origin https://huggingface.co/YOUR_USERNAME/aurora-one-gguf
git add .
git commit -m "Publish Aurora One GGUF"
git push origin main