Gemma 4 Maritime Medical — GGUF
A Gemma 4 2B model fine-tuned on the World Health Organization International Medical Guide for Ships, 3rd Edition (WHO IMGS), quantized to Q4_K_M for offline inference via Ollama or llama.cpp.
Built for
Vessel Ops AI — an offline-first medical and engineering assistant for vessels operating in deep-water environments where there is no doctor and no internet.
What this model is for
Maritime medical decision support at sea, where the Medical Person In Charge (MPIC) is a crew member — not a physician — and there is no shoreside telemedicine available. The model is fine-tuned to speak fluently in the dosages, drug names, and protocol structures used by the WHO IMGS, which is the de-facto standard medical reference carried by commercial vessels.
It is intentionally specialised. It is not a general medical assistant. Engine repairs, regulations, navigation, and general health questions should use the vanilla Gemma 4 base model instead — the Vessel Ops AI installer pulls both and routes traffic appropriately.
Quick start — Ollama
1ollama pull hf.co/nswitzer/gemma4-maritime-medical-GGUF
2ollama run hf.co/nswitzer/gemma4-maritime-medical-GGUF
Then ask:
A crew member has a deep laceration on the forearm with arterial bleeding. What should the MPIC do?
For best results, wrap calls in a retrieval-augmented prompt that injects the relevant WHO IMGS passages with page citations — that is what Vessel Ops AI does on every medical query, and it dramatically reduces hallucination on specific dosages.
Training
| |
|---|
| Base model | unsloth/gemma-4-e2b-it-unsloth-bnb-4bit |
| Method | QLoRA (rank 16, RSLoRA scaling) via Unsloth |
| Dataset | ~1,400 clinical Q&A pairs generated from the 938-chunk WHO IMGS index by Gemini, in ShareGPT format |
| Hardware | Free Kaggle T4 (16 GB VRAM) |
| Time | ~3 hours, 3 epochs, cosine LR schedule |
| Quantization | Q4_K_M GGUF (~1.3 GB) via model.push_to_hub_gguf |
The Unsloth training pipeline is what made this fine-tune feasible on free hardware — ~2× faster than vanilla HuggingFace Trainer with ~70% less VRAM. A shipping company could re-run the same notebook on their own fleet's incident history in an afternoon and get a deployable Ollama model out the other end.
Hardware requirements (inference)
| Laptop RAM | Expected response time |
|---|
| 8 GB | 10–15 seconds per response |
| 16 GB | 5–10 seconds per response |
| 32 GB+ | 3–5 seconds, or move to gemma4:e4b for stronger reasoning |
CPU-only inference works fine — no GPU required. This is the entire point of the GGUF/Ollama stack: usable AI on whatever laptop is already in the wheelhouse.
Limitations and disclaimer
This is an AI assistant intended as a second opinion when no physician is reachable. It is not a substitute for medical training or for shoreside telemedicine (TMAS). Every Vessel Ops AI response carries this disclaimer:
AI-generated guidance. Verify against physical manuals. Contact rescue services if situation is life-threatening.
The model:
- Inherits all biases and limitations of the base Gemma 4 2B model
- Has been specialised on the WHO IMGS specifically — it will be less reliable on medical content outside that corpus
- Should always be paired with retrieval over the source manual at inference time. The fine-tune helps with vocabulary and structure; RAG is what keeps the page-cited dosages honest.
Citation
The training corpus is derived from:
World Health Organization. International Medical Guide for Ships, 3rd Edition. WHO Press, 2007.
Please consult the original WHO publication for authoritative medical guidance.
Links
Ollama
An Ollama Modelfile is included for easy deployment.
This was trained 2x faster with
Unsloth
