Aurora Proelia is a compact 207M-parameter English language model from North ML. It is designed for lightweight local inference, short conversations, and concise explanations on CPU, Apple Silicon, or CUDA.

This release uses the original Proelia v9 checkpoint, the strongest preserved checkpoint from the project’s earlier chat experiments. It is packaged for the standard Hugging Face Transformers Auto* API.

What it is good at

Short conversational replies
Identity and introduction questions
Familiar facts and simple explanations
Lightweight local experimentation
Retrieval-augmented applications that provide source text in the prompt

Example responses from the checkpoint:

What is Python? Python is a general-purpose programming language known for readable syntax and a large ecosystem.

Explain photosynthesis in one sentence. Photosynthesis is how plants use light to make chemical energy from water and carbon dioxide.

What it is not

Aurora Proelia is not a frontier model, web browser, search engine, calculator, or autonomous tool-use agent. It does not know current events and cannot verify facts by itself. It may make mistakes on arithmetic, specialized subjects, multi-step reasoning, and broad science questions.

For current or specialized questions, an application should search first, select reliable sources, and pass the checked source text to the model. The application should validate the final answer before displaying it.

Run with Transformers

pip install torch transformers safetensors

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "North-ML1/Aurora-Proelia"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    trust_remote_code=True,
    dtype=torch.float32,
).eval()

messages = [{"role": "user", "content": "What is Python?"}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_tensors="pt",
    return_dict=True,
)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=64,
        do_sample=False,
        use_cache=False,
        pad_token_id=tokenizer.eos_token_id,
    )

prompt_tokens = inputs["input_ids"].shape[-1]
print(tokenizer.decode(output[0, prompt_tokens:], skip_special_tokens=True))

The repository includes the custom Aurora architecture files required by trust_remote_code=True. The tokenizer uses the checkpoint’s original Question: ... Answer: training format behind the normal chat API.

Evaluation snapshot

These are small engineering checks, not official leaderboard results:

Check	Result
Original v9 curated chat gate	14/14
Direct 20-prompt capability probe	15/20

The direct probe covered identity, short explanations, familiar facts, Python, photosynthesis, transformers, capitals, Earth, reinforcement learning, and basic arithmetic. The model was strongest on concise language and familiar knowledge, and remained unreliable on exact arithmetic and open-ended science. See BENCHMARKS.md for the test notes.

Model details

Property	Value
Parameters	206,942,208
Architecture	Aurora causal language model
Vocabulary	16,000 tokens
Context length	2,048 tokens
Recommended decoding	Greedy decoding for reproducible output
Intended hardware	CPU, Apple Silicon, or CUDA

Intended use

Use Aurora Proelia for research, local assistants, model experiments, and as a small component inside a retrieval or tool-use system. Keep search, source selection, arithmetic checks, safety filtering, and answer validation in the surrounding application.

License

This is a public North ML research release. No open-source license is granted by this repository; licensing and redistribution rights are reserved by North ML.