medgemma-maternal-lora-v2
a multilingual LoRA adapter for maternal health
2 min readupdated
medgemma-maternal-lora-v2 is the second release of a QLoRA adapter (a compact add-on for fine-tuning language models with minimal VRAM usage) built on the medical model google/medgemma-4b-it by Dawa Health with support from UNICEF. It answers questions about pregnancy, reproductive health, and contraception across seven languages spoken in Zimbabwe and Zambia: English, Shona, Ndebele, Tonga, Chinyanja, Bemba, and Lozi. Unlike v1, this version adds clinician-verified Q&A pairs extracted from documents, a foundational contraception dataset, and real user conversations from a chatbot. It is built for researchers and developers creating localized healthcare chat assistants.
Key features
- Training continued directly on top of the v1 adapter without adding new LoRA layers, so v2 attaches to the base model the exact same way as the first release.
- Trained exclusively on text data—the visual capabilities of the multimodal MedGemma 4B base model were neither touched nor tested.
- Handles natural code-switching within a single prompt, matching how users in the region actually communicate.
- Released under the open AGPL-3.0 license, though usage also requires complying with the MedGemma terms of use (Health AI Developer Foundations).
What you can do
- Integrate it into a support chatbot prototype (for platforms like WhatsApp) paired with mandatory clinician review and RAG over verified reference guides.
- Generate informational responses on pregnancy milestones, fertility, and cervical cancer prevention in low-resource African languages.
- Use it to benchmark medical LLM performance on low-resource languages and study text repetition loops during greedy decoding.
Example
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
"dawahealth/medgemma-maternal-lora-v2",
load_in_4bit=True,
)
Read more
DetailsMetrics: on a 48-question benchmark, the average chrF score jumped from 15.6 (base model) and 31.2 (v1) to 42.3; on an expanded 133-question set, the average chrF reached 34.9.
Training: QLoRA (rank 16, alpha 16, 4-bit base), taking roughly 10 GPU hours on a single NVIDIA T4.
Limitations: setting
repetition_penalty=1.15 during inference is recommended, as the model is prone to repeating phrases (especially in Bemba and Shona).Disclaimer: this is not a medical device, is unsuitable for diagnosis, and can confidently generate clinically inaccurate information—professional oversight is required.










Comments