NaviMed-UMB

📊 Live dashboard → NaviMed-UMB — interactive hardware-envelope table, model cards, and methodology.

Reproducible benchmark suite, model-card releases, and methodology for deploying Polish clinical large-language-model inference on a consumer-grade AMD RDNA 4 workstation (2× Radeon AI PRO R9700, gfx1201) — under ROCm 7.2 and vLLM 0.19.

The project began as a workstation engineering log and has grown into the infrastructure underlying a series of forthcoming methodology, quantization, and clinical-AI papers. It is the L1 pillar of a three-layer architecture (workstation → retrieval → arena) targeted at sovereign, on-premise medical-AI deployment.

Operating principle: scientific rigor over time and cost. Benchmarks use the full Tier-A statistical protocol (n = 10 per (quant, TP, N) cell, results reported as descriptive statistics — see METHODOLOGY.md §7.4). Where methodological faithfulness and wall-clock / GPU-hours conflict, rigor wins: e.g. the Llama-PLLuM-70B family is re-run at full n = 10 rather than shortcut.

Security note. This repository documents a workstation build, public benchmarks, and external release pipelines. Operational details — hostnames, domains, network addresses, secrets, patient data — are deliberately omitted; every artifact in this repo passes through pre-commit detect-secrets and a per-artifact embargo classification (METHODOLOGY.md §11).

What's inside

Directory	What it holds
bom/	Hardware bill of materials, power topology, UPS integration, PCIe map
environment/	Dated software/system manifests (pip freeze, ROCm, kernel) for reproducibility, plus sanity-test JSONs and coherence-probe outputs per release campaign
benchmarks/	vLLM / ROCm benchmark harness (Phase 1 envelope + Phase 2 scaling sweep), methodology, plotting scripts, and per-model results (benchmarks/results/ is gitignored — embargo-protected per §11)
scripts/	Orchestration scripts: sanity_sweep_pllum70b_awq.sh (three-stage Gate 1/2/3 runner), kill_port.sh (process-isolated cleanup), _env.sh (gfx1201 env vars)
calibration/	Polish clinical SmPC corpus for AWQ quantization (separately licensed; published on HuggingFace as a reusable artifact) + quantization scripts
ai-workstation-dashboard/	Real-time CPU/GPU monitoring (FastAPI + psutil + rocm-smi)
docs/sessions/	Long-form engineering session reports and debugging logs (one per substantive session)
logbook/	Short chronological build-diary entries (one per day of substantive activity)
history/	Layer 1 raw session logs imported from external assistants (Claude, GPT) — append-only
literature/	Source-of-truth references for papers and tools cited in the project
logs/	Per-run sweep logs (gitignored; embargo-protected)

Key documents

• METHODOLOGY.md — universal Phase 1 envelope + Phase 2 scaling protocol, the §8 vehicle-integrity boundary, and the §11.1/§11.2/§11.3 embargo classification applied across the 21-model suite.
• RELEASES.md — per-release notes (v0.1.0 → v0.5.0) and between-version events (the 2026-05-23 public Llama-PLLuM-70B AWQ release and the 2026-05-26 Run-3 consumer-GPU 8B/12B AWQ release live here).
• CITATION.cff — canonical citation (or use GitHub's "Cite this repository").
• AI_USAGE_DISCLOSURE.md — disclosure of generative-AI assistance, including the locally-served Bielik-11B that proofread documentation in May 2026.

Platform at a glance

AMD Ryzen 9 9950X3D · 2× GIGABYTE Radeon AI PRO R9700 32 GB (gfx1201, RDNA 4) · 96 GB DDR5-6000 · Kubuntu 24.04, kernel 6.17, ROCm 7.2.0 · vLLM 0.19.0. Full BOM and power topology: bom/readme.md. Exact software manifests: environment/.

Public artifacts

Released as part of (or alongside) this repository:

• Ten HuggingFace model cards — the Llama-PLLuM-70B family (mozarcik/Llama-PLLuM-70B-{base,instruct,chat}-{2412,2508,2512}-awq, 8 variants, 2× R9700 / TP=2 deployment target), plus two Run-3 consumer-GPU variants released 2026-05-26 fitting on a single R9700: mozarcik/Llama-PLLuM-8B-chat-2512-awq (Llama 3.1 base) and mozarcik/PLLuM-12B-chat-2512-awq (Mistral-Nemo base). To the author's knowledge the first public AWQ W4A16 / vLLM-native quantization of each variant. huggingface.co/mozarcik.
• One HuggingFace dataset — mozarcik/clinical-pl-smpc-awq-calibration, 418 fragments of Polish Charakterystyka Produktu Leczniczego (ChPL / SmPC) text from EMA, ~512 tokens each, covering 61 medicines drawn from a curated catalog of 81 INN across 9 NFZ drug programmes (pulmonology + thoracic oncology focus; No PHI). Reusable as a calibration corpus for other Polish-clinical quantization work.
• One LinkedIn announcement — activity 7464059097575907328.
• One Zenodo deposit — concept DOI 10.5281/zenodo.19851346 (auto-resolves to latest version; current version v0.5.0, version DOI minted at release; previous v0.4.0 at 10.5281/zenodo.20364953).

Status

Current release: v0.5.0 (2026-07-05) — Tier-A n=10 statistical re-runs of the Phase 2 suite + precision-ablation sub-study (raw compute complete, aggregation in progress) + N=1 single-stream anchor across all families + firmware provenance reconciliation + Layer 2 retrieval-architecture site documentation. Full release history and per-version highlights: RELEASES.md.

Most recent activity (2026-07-05): Tier-A n=10 re-runs on the standard METHODOLOGY §6 N grid {10, 25, 50, 100, 200, 500, 1000}, an N=1 single-stream anchor across all model families (incl. the eight 70B AWQ variants), and a same-checkpoint BF16↔AWQ precision-ablation sub-study. Per-N throughput/latency/power numbers are embargoed §11.2/§11.3 pending paper acceptance; engineering envelope, methodology, and site documentation are public.

AI assistance

Developed with generative AI tools — Claude Opus 4.7 and 4.8 (web and Claude Code CLI, orchestration and analysis), Claude Fable 5 (Claude Code subagents — site design, drafting, and review), GPT-5.5 Deep Thinking (web), Gemini (web review), and a locally-served Bielik-11B-v3.0-instruct-AWQ (via vLLM on one R9700, used for Polish-language proofreading of documentation in May 2026). All experimental design, hardware configuration, empirical measurements, and scientific claims are the sole responsibility of the author. Full table: AI_USAGE_DISCLOSURE.md.

License

This repository uses dual licensing:

• Code — MIT License.
• Documentation, methodology, lab logs, benchmark findings — CC BY 4.0.
• The calibration/ dataset — derived from third-party regulatory documents (EMA-published SmPC); governed by calibration/LICENSE, not the root CC-BY-4.0 / MIT licenses.

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

---

Reproducibility ethic: every piece of work here aims to either pass a pre-commit gate, leave an audit trail in logbook/ and docs/sessions/, or be flagged with an explicit embargo classification. For a non-serious take on the LLM benchmarking landscape, see benchmarks/assets/battle_of_LLM_models_gemini.png.