Aurora Proelia ChatML
Aurora Proelia ChatML is a 207M-parameter experimental variant of Aurora Proelia. It was SFT-trained on a conventional role-based ChatML surface so applications can send system, user, and assistant turns in a familiar format.
This is a separate candidate. The original Aurora-Proelia repository remains the native Question: / Answer: release.
ChatML format
Use this format for inference:
<|im_start|>system
You are Ember Proelia. Answer directly and concisely.<|im_end|>
<|im_start|>user
What is Python?<|im_end|>
<|im_start|>assistant
The model is a custom Aurora checkpoint. The included native Aurora runtime is the simplest path; a Transformers remote-code adapter is also provided below for normal Hub-style testing.
Transformers / Hugging Face test
The repository also includes a Transformers remote-code adapter, so it can be loaded through the normal AutoTokenizer and AutoModelForCausalLM APIs:
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "North-ML1/Aurora-Proelia-ChatML"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)
messages = [{"role": "user", "content": "What is Python?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=96, do_sample=False, use_cache=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
trust_remote_code=True is required because Aurora is a custom architecture; inspect the repository code before enabling it in an untrusted environment.
What changed
The checkpoint started from the released Aurora candidate and received 2,048 effective ChatML SFT updates over the existing answer-masked ChatML corpus. The pass was intended to teach the input/output surface, not to create a new general-knowledge model.
Evaluation
On a matched public benchmark mini-slice, the ChatML candidate changed as follows:
| Benchmark | Released Aurora | ChatML candidate |
|---|---|---|
| MMLU · 57 questions | 14/57 · 24.6% | 16/57 · 28.1% |
| ARC-Challenge · 50 questions | 13/50 · 26.0% | 15/50 · 30.0% |
| HellaSwag · 50 questions | 19/50 · 38.0% | 19/50 · 38.0% |
| GSM8K · 50 questions | 1/50 · 2.0% | 0/50 · 0.0% |
The exact runs are in benchmarks.json, regression_comparison.json, and chatml_smoke.json. These are transparent slices of public Hugging Face datasets, not official leaderboard evaluations.
The practical result is clearer than the small score changes: the ChatML candidate answers ordinary identity and Python prompts through the role-based format, while the released checkpoint often echoes the ChatML prompt. Arithmetic and uncertainty handling remain weak.
Limitations
This remains a small research model. It is unreliable for multi-step arithmetic, deep reasoning, current facts, specialized questions without context, and complex instruction following. Verify important answers and provide retrieval context when freshness or factual accuracy matters.
Local inference
pip install -r requirements.txt
python inference.py --prompt "What is Python?"
Omit --prompt to start an interactive chat:
python inference.py
Distribution
This is a public North ML research release. No open-source license is granted; licensing is reserved by the repository owner.
text-generation · aurora-proelia · chatml · north-ml · 207m