LumiSense

Public implementation of the LumiSense edge diagnosis framework.

Note

This repository is a streamlined, code-focused reorganization of the implementation accompanying the paper. It is intended to make the core workflow and logic easy to inspect. Because files were selectively consolidated for release, minor omissions or mistakes may remain. Please treat it as a reference implementation and report issues when you find them.

LumiSense adapts a small language model for evidence-grounded industrial IoT root-cause diagnosis. The code is organized by workflow stage:

Path	Purpose
code/sft/	Supervised fine-tuning data preparation, training, and evaluation.
code/kd/	Verified teacher annotation and class-token distillation.
code/wise_ft/	LoRA adapter interpolation between SFT and KD checkpoints.
code/quantization/	LoRA merge, GGUF export, quantization gate, and CPU benchmark helpers.
code/cascade/	Confidence-based edge-cloud routing utilities.
code/baselines/	Traditional and prompted-LLM baseline scripts.

A small fixture under tests/fixtures/mini_cares/ is provided only for checking data contracts and command wiring.

Released LumiSense model weights are hosted under the author's Hugging Face account.

Install

python3 -m pip install -e ".[test]"

For Apple-Silicon MLX experiments:

python3 -m pip install -e ".[mac,test]"

Traditional baselines additionally use scikit-learn:

python3 -m pip install -e ".[baseline]"

Smoke Tests

python3 -m unittest discover -s tests -v
DRY_RUN=1 code/sft/run_sft.sh --model /path/to/Qwen3.5-0.8B-MLX-4bit

All stage scripts accept --data-dir or DATA_DIR=... so the fixture can be replaced by a local dataset that follows the expected format.

Citation

If you use the LumiSense code or released model, please cite the associated paper. The final bibliographic record will be added after publication:

% TODO: replace the placeholder fields with the final publication metadata.
@inproceedings{lumisense2026,
  title     = {LumiSense: Evidence-Grounded Edge Sensor Diagnosis with Adapted Small Language Models},
  author    = {He, Sheng},
  booktitle = {Publication venue to be added},
  year      = {2026}
}

License

MIT. Model weights, adapters, raw prediction outputs, and paper-run data are not included in this repository.