Sakhi – MedGemma 1.5 4B Maternal Health LoRA
This repository contains a LoRA adapter fine-tuned on maternal, antenatal, postpartum, and newborn health data for ASHA worker clinical assistance.
Base model: google/medgemma-1.5-4b-it
Training Method
- Base model loaded in 4-bit NF4 quantization (QLoRA)
- LoRA rank: 16
- LoRA alpha: 16
- Optimizer: paged_adamw_8bit
- Learning rate: 2e-4
- Epochs: 1
- Max sequence length: 1024
- Hardware: Kaggle 2×T4 GPUs
Datasets:
- lavita/ChatDoctor-HealthCareMagic-100k (filtered maternal/newborn)
- medalpaca/medical_meadow_wikidoc_patient_information (OB/GYN filtered)
Intended Use
This adapter is designed to improve maternal and neonatal health reasoning for:
- Antenatal care interpretation
- High-risk pregnancy identification
- Neonatal danger sign detection
- Breastfeeding and postpartum support
- India-specific maternal health contexts (IFA, anaemia, PHC referral)
It is intended for use inside Sakhi — an AI clinical assistant for ASHA workers.
Prohibited Use
This model must NOT be used for:
- Autonomous medical diagnosis
- Replacing licensed medical professionals
- Emergency decision-making without human oversight
- Direct-to-patient unsupervised deployment
Safety Limitations
- Model may hallucinate.
- Does not replace trained clinicians.
- Should be used only as decision-support.
- Must recommend PHC referral when uncertainty exists.
System Prompts
CHECKUP Mode
Respond strictly in JSON format:
{{
"risk_level": "green" | "yellow" | "red",
"risk_reason": "...",
"what_sakhi_noticed": ["...", "..."],
"what_to_tell_patient": "...",
"what_to_do_next": "...",
"follow_up_date": "YYYY-MM-DD or null"
}}
Red flag example:
BP > 140/90 → risk_level: red
NEWBORN Mode
Age-specific rules:
- Jaundice after Day 14 → red
- Weight loss after Day 7 → yellow/red
- Cord redness or smell → red
- Breathing > 60/min → red
Same JSON schema as CHECKUP mode.
CHAT Mode
Conversational.
Always address the ASHA worker directly.
Use simple, clear English.
Recommend PHC referral when unsure.
License
