Modelpublic

Fernandosr85/parteirabr-maternal-triage-adapter

sourceHugging Facellama3.3updated 3mo agoView on Hugging Face

0likes

Model Card

banner

ParteiraBR — Maternal Health Triage Adapter

LoRA adapter for maternal health triage in Brazilian Portuguese — designed for riverine, quilombola, and backland communities where traditional midwives are often the first point of care and formal healthcare services are distant.

Fine-tuned on Llama-3.3-70B-Instruct via Adaption's AutoScientist platform.

Covers 13 structured clinical risk signals across 3 urgency tiers (vermelho / amarelo / verde), grounded in official Brazilian Ministry of Health publications — all responses in PT-BR.

Evaluation results

Training Winrates

The adapted model outperforms the base model with 80% win rate vs 21% on held-out medical evaluation — a +281% relative improvement over the base model.

Model	Win Rate
Base (`Llama-3.3-70B-Instruct`)	21%
Adapted (`parteirabr-adapter`)	80%

Train/Eval Metrics

Metric	Value
Initial train loss	~1.59
Final validation loss	~0.42
Training steps	163
LR scheduler	cosine (warmup)
Gradient norm	spike → stable

Train loss converged steadily over 163 steps. Validation loss tracked closely, confirming generalization without overfitting. Learning rate followed cosine schedule with warmup. Gradient norm stabilized after initial spike, indicating stable optimization throughout.

Dataset quality

Metric	Value
Dataset grade	A
Score before / after	9.0 → 10.0
Quality improvement	+11.1% (Adaption Adaptive Data remastering)
Medical domain percentile	43.9 → 57.7
Total examples	1,200 → 28,700 (augmented)

Model details

Field	Value
Base model	`meta-llama/Llama-3.3-70B-Instruct` (70B)
Trained model name	`adaption_llama_3_3_70b_instru_parteira_br_maternal_gui_80fa5adc`
Training method	Supervised Fine-Tuning (SFT) + LoRA
Data format	Chat (instruction/completion)
Domain	Brazilian maternal health triage — PT-BR
Category	Healthcare

Training dataset

[fernandosr85/parteirabr-signalgrounding](https://www.kaggle.com/datasets/fernandosr85/parteirabr-signalgrounding)

1,200 instruction/completion pairs across 13 clinical signals:

Tier	Signal	Examples
🔴 Vermelho	Sangramento vaginal	50
🔴 Vermelho	Dor de cabeça forte com alterações visuais	50
🔴 Vermelho	Convulsão	50
🔴 Vermelho	Diminuição ou ausência de movimentos do bebê	50
🔴 Vermelho	Perda de líquido amniótico antes do termo	50
🔴 Vermelho	Febre alta no pós-parto com mau cheiro	50
🟡 Amarelo	Inchaço súbito de rosto e mãos	100
🟡 Amarelo	Ardência ou dor para urinar	100
🟡 Amarelo	Febre baixa persistente	100
🟡 Amarelo	Ausência de acompanhamento de pré-natal	100
🟢 Verde	Enjoo e náusea no primeiro trimestre	200
🟢 Verde	Cansaço e sonolência	200
🟢 Verde	Dúvidas sobre amamentação	200

Quality controls

8-dimension rubric (16 pts max): correct referral, welcoming tone, simple language, respect for traditional knowledge, no risk minimization, no prescription or diagnosis, concrete referral pathway, structural clarity. Dataset mean: 15.41/16.

3-tier safety gate: blocks home remedies for emergencies, wait-and-see language for red signals, diagnosis, and prescription. All 1,200 examples passed with safety_flag: ok.

Grounding

All 13 signals traced to specific pages in official MoH publications:

—Caderneta Brasileira da Gestante (2026)
—Estratificação de Risco Gestacional — SES-MG (2025)
—Atenção ao Pré-Natal de Baixo Risco — CAB nº 32 (2012)
—Saúde da Criança: Aleitamento Materno — CAB nº 23 (2015)
—Manual de Gestação de Alto Risco (2022)

Safety Principle

The core constraint of this adapter: when any risk signal is present, always refer to formal healthcare services — immediately for red signals, within 24–48h for yellow signals.

	Example
✅ Correto	"Isso é uma emergência. Procure o SAMU 192 AGORA."
❌ Bloqueado	"Tome um chá e veja se melhora até amanhã."

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE_MODEL = "meta-llama/Llama-3.3-70B-Instruct"
ADAPTER_PATH = (
    "/kaggle/input/parteirabr-maternalhealth-adapter/"
    "other/llama-3-3-70b-instruct-lora/1"
)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    device_map="auto",
    torch_dtype="auto",
)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_PATH,
)

prompt = """ParteiraBR — Guia de Orientação em Saúde Materna Tradicional

Você é um assistente de acolhimento e navegação em saúde materna.
Diante de qualquer sinal de risco, encaminhe imediatamente ao serviço
de saúde. Nunca diagnostique, prescreva ou sugira remédios caseiros
para emergências.

Contexto: uma gestante de uma comunidade ribeirinha no Amazonas
perguntou:
Estou grávida e tive um sangramento forte agora. Estou com medo.
O que eu faço?"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
)

generated_tokens = outputs[0][inputs["input_ids"].shape[1]:]
print(tokenizer.decode(generated_tokens, skip_special_tokens=True))

Model repositories

Platform	Link
HuggingFace Model	Fernandosr85/parteirabr-adapter
Kaggle Model	fernandosr85/parteirabr-maternalhealth-adapter
Kaggle Dataset	fernandosr85/parteirabr-signalgrounding

Credits

—Fine-tuning platform: Adaption — AutoScientist & Adaptive Data
—Challenge: AutoScientist Challenge 2026
—Training infrastructure: Adaption compute credits
—Dataset remastering: Adaption Adaptive Data pipeline (Grade A, +11.1% improvement)
—Grounding validator: ParteiraBR
—Author: Fernando Rodrigues · Kaggle: fernandosr85 · HuggingFace: Fernandosr85

Disclaimer

Experimental research artifact submitted to the AutoScientist Challenge 2026 (Healthcare category). The adapter does not replace evaluation by a qualified healthcare professional. All clinical guidance generated by the model must be reviewed before use in real community health settings. Traditional midwifery knowledge is respected and valued — this adapter is designed to complement, not replace, community health networks.