Repository navigation
Sofia Distilled v0.1.0
This release provides actual trained weights and reproducible training for two Sofia research models using explicit classical controls inspired by Quantum Artificial Intelligence with Verifiable Kernels.
- Sofia Edge: 325-parameter NumPy vibration classifier distilled from a verified product-kernel ridge teacher. Synthetic held-out accuracy is 99.933%; the supervised-only control reaches 100%. There is no measured-machine validation or claimed distillation quality advantage.
- Sofia Chat: an 18-layer, 404,558,464-parameter Qwen-derived checkpoint, with 18.11% fewer parameters than its teacher. Its training combines teacher-token distillation and product-kernel representation alignment. Actual held-out generations contain factual errors; the checkpoint is for research and auditing, not dependable engineering advice.
The repository includes original synthetic data, original AI-assisted English Chat references, split specifications, model cards, measured baselines, retained test generations, pinned teacher provenance, and SHA-256 manifests. Chat inherits Qwen pretraining; no foundation pretraining from random initialization is claimed. Full Chat weights are distributed through Hugging Face.
Validation: 25 local tests passed; scientific-core CI passed. The same-environment independent verification rerun reproduced all 810 stored arrays exactly. The product kernel has an efficient exact classical equivalent. No quantum hardware execution or quantum computational advantage is claimed.
See the English README for installation, inference, combined Edge/Chat use, training, and public-release verification commands. Apache-2.0; upstream Qwen attribution is retained.