Instructions to use docvm/sakhi-medgemma-1.5-4b-maternal-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use docvm/sakhi-medgemma-1.5-4b-maternal-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="docvm/sakhi-medgemma-1.5-4b-maternal-GGUF",
	filename="sakhi-medgemma-maternal-Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use docvm/sakhi-medgemma-1.5-4b-maternal-GGUF with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M

Use Docker

docker model run hf.co/docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use docvm/sakhi-medgemma-1.5-4b-maternal-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "docvm/sakhi-medgemma-1.5-4b-maternal-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "docvm/sakhi-medgemma-1.5-4b-maternal-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M

Ollama
How to use docvm/sakhi-medgemma-1.5-4b-maternal-GGUF with Ollama:
```
ollama run hf.co/docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M
```

Unsloth Studio

How to use docvm/sakhi-medgemma-1.5-4b-maternal-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for docvm/sakhi-medgemma-1.5-4b-maternal-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for docvm/sakhi-medgemma-1.5-4b-maternal-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for docvm/sakhi-medgemma-1.5-4b-maternal-GGUF to start chatting

Atomic Chat new
Docker Model Runner
How to use docvm/sakhi-medgemma-1.5-4b-maternal-GGUF with Docker Model Runner:
```
docker model run hf.co/docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M
```

Lemonade

How to use docvm/sakhi-medgemma-1.5-4b-maternal-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.sakhi-medgemma-1.5-4b-maternal-GGUF-Q4_K_M

List all available models

lemonade list

sakhi-medgemma-1.5-4b-maternal-GGUF

A QLoRA fine-tuned, merged, and Q4_K_M-quantized version of google/medgemma-1.5-4b-it, specialized for maternal and neonatal clinical triage in the Indian rural health context. This is the active production model powering Sakhi — an AI clinical companion for ASHA (Accredited Social Health Activist) workers.

The LoRA adapter (docvm/sakhi-medgemma-1.5-4b-maternal) was merged into the base model weights, converted to GGUF via llama.cpp, and quantized to Q4_K_M (~2.5 GB). It runs on CPU via Ollama and exposes an OpenAI-compatible endpoint.

Intended Use

This model is designed to assist ASHA workers — trained community health volunteers in rural India — during antenatal checkups and newborn postnatal visits. It is called by the Sakhi backend to:

Stratify maternal and neonatal risk (green / yellow / red)
Flag warning signs (hypertension, severe anaemia, cord complications, etc.)
Suggest referral decisions aligned with MOHFW/WHO guidelines
Answer free-form clinical questions in a field-appropriate tone

This model is a clinical decision support tool. It does not diagnose. All outputs should be reviewed by a trained health worker before any action is taken.

Training

Fine-tuning method

QLoRA (4-bit NF4 quantization of base weights during training) via Unsloth on Kaggle (2×T4).

Parameter	Value
LoRA rank	16
LoRA alpha	16
LoRA dropout	0.05
Target modules	all-linear
Optimizer	paged_adamw_8bit
Learning rate	2e-4
LR schedule	cosine
Epochs	1
Batch size	2 (grad accumulation steps = 4, effective batch = 8)
Max sequence length	512
Training time	~4.2 hours
Trainable parameters	38.5M / 4.34B (0.89%)
Final training loss	2.13

Training data

Two public HuggingFace datasets, filtered to maternal/neonatal content via keyword matching:

Dataset	HF repo	Filtered size
ChatDoctor-HealthCareMagic-100k	`lavita/ChatDoctor-HealthCareMagic-100k`	5,000 examples
WikiDoc Patient Information	`medalpaca/medical_meadow_wikidoc_patient_information`	1,500 examples

Filter keywords: pregnancy, antenatal, postpartum, newborn, neonate, breastfeed, jaundice, preeclampsia, gestational diabetes, anaemia, low birth weight, cord, lactation, miscarriage, ectopic, folic acid, iron.

Total after 95/5 train/eval split: ~5,300 train / ~280 eval.

Key training objectives

Indian clinical context: recognition of locally prevalent risk patterns (severe anaemia, eclampsia, low birth weight) common in Rajasthan and similar settings
Output reliability: improved JSON schema compliance for structured triage output, reducing post-processing failures in production

Quantization

The LoRA adapter was merged into the base google/medgemma-1.5-4b-it weights (bfloat16), then converted and quantized using llama.cpp:

python convert_hf_to_gguf.py merged_model/ --outtype bf16
./llama-quantize model-bf16.gguf model-Q4_K_M.gguf Q4_K_M

Final size: ~2.5 GB (Q4_K_M). Runs on CPU with ~6 GB RAM.

The merge and quantize pipeline is fully documented in model/merge-and-quantize.ipynb.

How to Use

Ollama (recommended)

ollama pull docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M
ollama run docvm/sakhi-medgemma-1.5-4b-maternal-GGUF:Q4_K_M

The model exposes an OpenAI-compatible endpoint at http://localhost:11434/v1.

llama.cpp

./llama-cli -m sakhi-medgemma-1.5-4b-maternal-Q4_K_M.gguf \
  --chat-template gemma \
  -p "You are Sakhi, an AI clinical companion for ASHA workers..."

Example prompt (Sakhi triage format)

You are a maternal triage AI.

Classify the case into exactly one of:
Triage: HIGH
Triage: MODERATE
Triage: LOW

Escalate to HIGH if any of the following are present:
- BP ≥160 systolic or ≥110 diastolic
- Seizures, convulsions
- Heavy bleeding
- Signs of sepsis (fever + rigors + abdominal tenderness postpartum)
- Visual disturbance + hypertension
Otherwise classify appropriately.

Output strictly in this format:
Triage: <HIGH/MODERATE/LOW>
Reason: <one short sentence>

Case: 26-year-old, 34 weeks pregnant. BP 162/108. Headache and visual
disturbance since morning. No bleeding.

Limitations

Trained on English-language medical Q&A data; Hindi-language performance is untested at the model level (the Sakhi app handles Hindi via prompt instruction).
Training data is filtered public datasets, not real patient records. Clinical thresholds were applied in post-processing and evaluation, not through supervised fine-tuning on labeled triage decisions.
The model is not a replacement for clinical judgment or specialist review.
Not validated in a prospective clinical setting.

Part of the Sakhi Project

Resource	Link
Sakhi app (live demo)	https://sakhi-asha.vercel.app
Backend API	https://docvm-sakhi-api.hf.space/health
LoRA adapter (pre-merge)	`docvm/sakhi-medgemma-1.5-4b-maternal`
GitHub repo	https://github.com/orcus108/sakhi
Fine-tuning notebook	`model/finetuning-medgemma.ipynb`
Merge + quantize notebook	`model/merge-and-quantize.ipynb`

Built for the Google MedGemma Impact Challenge · Kaggle · February 2026.

License

Health AI Developer Foundations Terms of Use

Downloads last month: 26

GGUF

Model size

4B params

Architecture

gemma3

Hardware compatibility

4-bit

Model tree for docvm/sakhi-medgemma-1.5-4b-maternal-GGUF

Base model

google/medgemma-1.5-4b-it

Quantized

(36)

this model

docvm
/

sakhi-medgemma-1.5-4b-maternal-GGUF

sakhi-medgemma-1.5-4b-maternal-GGUF

Intended Use

Training

Fine-tuning method

Training data

Key training objectives

Quantization

How to Use

Ollama (recommended)

llama.cpp

Example prompt (Sakhi triage format)

Limitations

Part of the Sakhi Project

License

Model tree for docvm/sakhi-medgemma-1.5-4b-maternal-GGUF

Space using docvm/sakhi-medgemma-1.5-4b-maternal-GGUF 1