Downloads · 30 days
0
itskios09/daktarai-gemma4-african-clinical-v2
daktarai-gemma4-african-clinical-v2 is a text generation model from itskios09. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
Downloads · 30 days
0
Access
Public
Updated Apr 5, 2026
Repo size
326 MB
Likes
0
Public
Click a slice to open those files.
.safetensors294 MB · 90%
From the Hugging Face model README
Daktari (Swahili) — Doctor
DaktariAI is the first large language model fine-tuned on real, anonymized Sub-Saharan African electronic health record data. Designed for community health workers (CHWs) operating in low-resource, offline environments.
v2 upgrades over v1: Larger LoRA rank (r=32), longer sequence length (1024), 5 training epochs, trained on RTX 6000 Ada (48GB VRAM).
3.5 billion people live without access to a doctor. In Sub-Saharan Africa, community health workers make life-or-death clinical decisions daily with no decision support, no specialist backup, and no internet connectivity.
Existing medical AI is trained on USMLE exams and Indian medical boards — reflecting Western disease burdens. No model has been fine-tuned on real Sub-Saharan African EHR data. Until now.
| Property | V1 | V2 |
|---|---|---|
| Base model | Gemma 4 E4B | Gemma 4 E4B |
| Fine-tuning method | Unsloth + LoRA | Unsloth + LoRA |
| LoRA rank | r=16 | r=32 |
| Max sequence length | 512 | 1024 |
| Training epochs | 2 | 5 |
| Training loss | 1.3326 | 1.1120 |
| Validation loss | 1.1915 | 1.2338 |
| Trainable parameters | 36.7M (0.60%) | 73.4M (1.20%) |
| Training hardware | NVIDIA L4 22GB | NVIDIA RTX 6000 Ada 48GB |
| Training time | 227 min | 244 min |
Real anonymized clinical encounters from a production African EHR deployment:
Anonymization pipeline:
General medical knowledge from openlifescienceai/medmcqa — 194K AIIMS & NEET PG exam questions.
Total training pairs: 15,464 (41.8% African clinical, 58.2% MedMCQA)
| GPT-4 / Gemini | Existing Med AI | DaktariAI v2 | |
|---|---|---|---|
| Training data | General internet | USMLE / India exams | Real African EHR |
| African disease context | Poor | Poor | Native |
| Offline capable | No | No | Yes |
| CHW-optimized | No | No | Yes |
| Drug formulary | Western | Western/Indian | African (TDF+3TC+DTG, etc.) |
| Disclaimers | Excessive | Excessive | Minimal — actionable output |
Case 1 — HIV + TB + Pregnancy (Triage)
Input: F patient aged 25-34, HIV Stage 2, poor ART adherence, pregnant, on TB treatment.
DaktariAI v2: Likely Diagnosis: HIV-1 Infection Stage 2. ART: Initiate TLD 200mg/300mg/100mg. Integrate ART into routine treatment to improve adherence. TB: Continue TB treatment as prescribed. Pregnancy: Ensure antenatal care and PMTCT prophylaxis. Weight: Underweight — counsel on nutrition.
Case 2 — Uncontrolled DM + Hypertension
Input: M patient aged 35-49, BP 158/98, Diabetes, FBS 14.2.
DaktariAI v2: Diagnosis: Hypertension, Type 2 Diabetes Mellitus. Treatment: Continue Metformin 500mg twice daily. Add Semaglutide. Continue antihypertensives.
Case 3 — Preeclampsia (MCH)
Input: F patient aged 18-24, 28 weeks gestation, BP 148/96, urine protein 2+, oedema.
DaktariAI v2: Diagnosis: Hypertension in Pregnancy. Treatment: Refer to facility for management of hypertension in pregnancy.
Case 3 demonstrates CHW-appropriate decision-making — immediate referral rather than academic analysis.
| Base Gemma 4 E4B | DaktariAI v2 | |
|---|---|---|
| Response style | Academic tables, verbose | Direct, actionable |
| Disclaimers | 3+ per response | Minimal |
| Drug specificity | Generic | TLD 200/300/100mg, Metformin 500mg |
| CHW usability | Too complex | Ready to act on |
| Referral decisions | Long analysis, no conclusion | Immediate clear decision |
# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh
# Run DaktariAI v2 offline
ollama run itskios09/daktarai-gemma4-african-clinical-v2
Requirements: 8GB RAM, 5GB storage, no internet after initial download.
@misc{daktarai2026,
author = {Kioko, Fredrick},
title = {DaktariAI v2: Gemma 4 Fine-tuned on Real African Clinical Data},
year = {2026},
publisher = {HuggingFace},
url = {https://huggingface.co/itskios09/daktarai-gemma4-african-clinical-v2}
}