Experiments - August 2026
Collection
3 items • Updated
Aurora Proelia
Aurora Proelia is a compact, identity-aligned language model from North ML. It is a 207M-parameter-class causal language model designed for lightweight local text generation. Its identity is Aurora Proelia.
This card describes the model’s observed behavior, not an aspirational benchmark result.
This model is a failed experiment. Use it for running on device quick assistants that don't need much. Fine-tune if it doesn't regress.
What it is good for
- Short identity and introduction prompts
- Short, familiar factual explanations
- Lightweight local completion on Apple Silicon, CUDA, or CPU
- Simple prompt experiments where low latency and small memory use matter
Known limitations
Aurora Proelia is not a frontier model and should not be treated as one. In testing, it is unreliable for:
- Multi-step arithmetic and exact calculation
- Deep reasoning or long explanations
- Open-ended creative writing
- Specialized technical questions without context
- Current events or facts that may have changed
- Consistently following complex instructions
It does not browse the web or call tools by itself. To answer current questions, an application must perform search or retrieval and pass the selected source text to the model. Retrieved sources still need to be checked by the application or user.
Hugging Face benchmark snapshot
These scores were run on the current public checkpoint with the native Aurora runtime on Apple Silicon. They use real Hugging Face test datasets, but they are transparent sampled runs rather than official full leaderboard evaluations.
| Benchmark | Hugging Face test data | Sample | Scoring | Result |
|---|---|---|---|---|
| MMLU | all/test |
285 questions, 5 per subject across 57 subjects | Four-choice next-token accuracy | 67/285 · 23.5% |
| GSM8K | main/test |
20 questions | Final-number exact match | 0/20 · 0.0% |
The MMLU sample is close to four-choice chance, and the GSM8K sample shows that this checkpoint is not reliable at multi-step arithmetic. Both runs used deterministic decoding with seed 20260815. These results should be read alongside the capability examples above, not as claims of broad reasoning ability.
Model details
| Parameters | 206,942,208 |
| Architecture | Aurora causal language model |
| Tokenizer | 16,000 tokens |
| Context length | 2,048 tokens |
| Runtime | Native Aurora runtime |
| Release | Public preview |
The weights started from the original Ember Proelia preview and received a small response-masked identity SFT pass. A conservative blend with the untouched parent was used to reduce behavioral drift. This training was intended to change the model’s identity, not to claim a general capability improvement.
No later repair checkpoint is published here: the local SFT repair candidates regressed behavior and were rejected after evaluation.
Local inference
pip install -r requirements.txt
python infer.py --checkpoint model.safetensors
For the most predictable output, use the native runtime with greedy decoding. Verify generated text before relying on it.
Intended use
Use Aurora Proelia for research, local experimentation, identity testing, and small text-generation applications. It is suitable as a compact component in a larger retrieval or tool-use system, but the surrounding application must provide search, validation, and safety controls.
License and distribution
This is a public North ML preview. Public visibility does not grant permission to redistribute the weights or publish derivatives. No open-source license is granted; licensing is reserved by the repository owner.
Tags
text-generation · aurora-proelia · north-ml · public-preview