You can deploy this model and start earning money today!

gthor7/
pcos-health-ai-pipeline

Name: pcos-health-ai-pipeline
Availability: InStock

Views

No views yet

PCOS Health AI Pipeline

A production-grade AI pipeline for a hormonal health tracking mobile application focused on Polycystic Ovary Syndrome (PCOS).

Architecture

User Input → Llama Guard (input safety) → PCOS LLM + RAG → Llama Guard (output safety) → Response
                                            ↓
                                     NeMo Guardrails (topic control, referral triggers)

Components

Module	File	Description
Dataset Pipeline	`pcos_health_ai/dataset_pipeline.py`	Curates, synthesizes, and preprocesses training data
LoRA Fine-Tuning	`pcos_health_ai/lora_finetuning.py`	QLoRA + SFT training configuration
Safety Guardrails	`pcos_health_ai/safety_guardrails.py`	Llama Guard + rule-based medical safety
RAG Pipeline	`pcos_health_ai/rag_pipeline.py`	ChromaDB vector store + retrieval-augmented generation
Config	`pcos_health_ai/config.py`	Centralized configuration and constants

Base Model Recommendation

Model	Pros	Cons	USMLE Score
BioMistral-7B ⭐	Medical pre-trained, AWQ available (4.68GB)	Slightly older base	~75%
Llama-3-8B-Instruct	Best instruction following, strong GGUF support	No medical pre-training	~70% (base)
MedGemma-4B-IT	Newest SOTA, multimodal ready	Gated access, newer ecosystem	~84%

Recommendation: BioMistral/BioMistral-7B for clinical accuracy, or meta-llama/Meta-Llama-3.1-8B-Instruct for best conversational quality.

Datasets Used

lavita/ChatDoctor-HealthCareMagic-100k — 100K real patient-doctor Q&A
ruslanmv/ai-medical-chatbot — 250K+ patient-doctor exchanges
medalpaca/medical_meadow_wikidoc_patient_information — Clinical knowledge
Amod/mental_health_counseling_conversations — Empathetic counseling
Malikeh1375/medical-question-answering-datasets — Unified medical QA bundle

Training Presets

Preset	Base Model	LoRA r	Epochs	Max Seq	Hardware
Quick Test	BioMistral-7B	8	1	512	a10g-small
Standard	BioMistral-7B	16	3	2048	a10g-large
High Quality	Llama-3.1-8B-Instruct	64	5	4096	a100-large
Maximum	Llama-3.1-8B-Instruct	256	8	4096	a100-large

Safety Guardrails (3 Layers)

Rule-Based — Fast, deterministic checks for medication prescriptions, emergency symptoms, diet culture
Llama Guard 3 — Model-based content classification with custom PCOS medical taxonomy (F1=0.945)
NeMo Guardrails — Programmable Colang dialogue flows for topic control

Quick Start

bash

pip install datasets transformers peft trl bitsandbytes accelerate sentence-transformers chromadb faker

# 1. Generate dataset
python -m pcos_health_ai.dataset_pipeline

# 2. Run safety guardrails demo
python -m pcos_health_ai.safety_guardrails

# 3. Run RAG pipeline demo
python -m pcos_health_ai.rag_pipeline

# 4. View training configuration
python -m pcos_health_ai.lora_finetuning

Research References

Med42 (arXiv:2404.14779) — LoRA all-linear targets, lr=1e-4
BioMistral (arXiv:2402.10373) — Medical pre-training, r=16, dropout=0.05
LoRA Without Regret (2025) — r=256 for full-FT parity
ChatDoctor (arXiv:2303.14070) — Patient-doctor conversation dataset
SoulChat (arXiv:2311.00273) — Empathetic health conversation patterns
DISC-MedLLM (arXiv:2308.14346) — Multi-stage medical SFT strategy
Llama Guard (arXiv:2312.06674) — Input/output safety classification
NeMo Guardrails (arXiv:2310.10501) — Programmable safety rails

License

This pipeline is for research and educational purposes. Always consult healthcare professionals for medical decisions.

gthor7/pcos-health-ai-pipeline