Cite this work

NaviMed-UMB is archived on Zenodo with a versioned DOI chain. It ships a CITATION.cff for GitHub's "Cite this repository" button.

DOIs

For most purposes, cite the concept DOI. It auto-resolves to the latest release. Cite a specific version DOI when reproducibility against an exact snapshot matters.

Canonical citation

From CITATION.cff (CFF 1.2.0):

Minarowski, Ł. (2026). NaviMed-UMB: hardware envelope studies for local AI deployment on consumer RDNA 4 GPUs (Version 0.4.0) [Software]. Zenodo. https://doi.org/10.5281/zenodo.19851346

BibTeX

@software{minarowski_navimed_umb_2026,
  author    = {Minarowski, {\L}ukasz},
  title     = {{NaviMed-UMB: hardware envelope studies for local
               AI deployment on consumer RDNA 4 GPUs}},
  year      = {2026},
  version   = {0.4.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19851346},
  url       = {https://github.com/kicrazom/navimed-umb},
  license   = {CC-BY-4.0}
}

Author

Łukasz Minarowski, MD, PhD
Department of Respiratory Physiopathology, Medical University of Białystok, Poland
ORCID: 0000-0002-2536-3508
Email: lukasz.minarowski@umb.edu.pl

Licensing

The repository uses dual licensing:

• Code: MIT.
• Documentation, methodology, lab logs, benchmark findings: CC BY 4.0.
• The calibration dataset: derived from third-party EMA-published regulatory documents (SmPC). It is governed by its own license, not the root CC-BY-4.0 / MIT.

Released model checkpoints carry the license of their base family. llama3.1 for the Llama-PLLuM derivatives, apache-2.0 for the Mistral-Nemo-derived 12B. See Models.

AI-usage disclosure

Transparency about AI assistance is a default policy for this project. Documentation prose, debugging dialogue, and sounding-board discussion used generative-AI tools: Claude (Anthropic, web + Claude Code CLI), GPT (OpenAI, web), and Gemini (Google, web review). A locally-served Bielik-11B-v3.0-instruct-AWQ (vLLM on one R9700) handled Polish-language editorial work.

All experimental design, hardware configuration, empirical measurements, and scientific claims are the sole responsibility of the author. AI tools did not execute experiments, did not access the workstation, did not generate quantized weights, and did not validate any benchmark measurement. The work has a single author. Consistent with COPE and ICMJE guidance, AI tools are documented as research aids, not co-authors. Full table and per-role detail: AI_USAGE_DISCLOSURE.md.

Acknowledgements

Quantization compute for the 70B family ran on an AMD Instinct MI300X through the AMD Developer Cloud, powered by DigitalOcean, provided as a promotional credit under the AMD Developer Program. The base models come from the PLLuM consortium (SpeakLeash, OPI-PIB, NASK, Wrocław University of Science and Technology) and the Polish Ministry of Digital Affairs, published via CYFRAGOVPL. The calibration corpus derives from SmPC text published by the European Medicines Agency.