Fernandosr85/healthbr-coverageguide-adapter

HealthBR-CoverageGuide — Brazilian Healthcare Coverage & Access Guidance Adapter

LoRA adapter fine-tuned on Llama-4-Scout-17B-16E-Instruct (109B) for safe, non-diagnostic Brazilian healthcare coverage and SUS access guidance, via Adaption's AutoScientist platform.

---

The problem this adapter addresses

Healthcare-adjacent LLM evaluations usually focus on diagnosis, which is exactly the wrong target for a patient-facing administrative assistant — diagnosis is a clinical act, not an administrative one. Given a raw citizen request like "meu plano disse que não cobre um exame que o médico pediu, isso pode?", a base model typically responds with an unqualified, potentially wrong answer:

"Não, isso não é coberto, o plano não é obrigado."

The safe institutional response follows an explicit structure — formal opening, objective context, explanatory body citing the correct regulatory pathway, a concrete next step, and a mandatory disclaimer:

"Prezado(a) cidadão(a), esclarecemos que os planos de saúde contratados a partir de janeiro de 1999 devem seguir o Rol de Procedimentos da ANS... Caso a operadora tenha negado a cobertura, recomenda-se solicitar a justificativa formal por escrito... Esta orientação tem caráter administrativo e não substitui avaliação médica ou análise jurídica individualizada."

This adapter teaches the model to apply an explicit HealthBR Guidance Guide (structure, tone, required vocabulary, prohibited claims) to a raw input, and to comply with a deterministic 14-point safety/quality rubric, gated by hard safety patterns that block dangerous completions before any other scoring.

---

Adaptive Data results

Metric	Before	After
Quality score	9.0	9.8
Quality grade	B	A
Relative improvement	—	+8.9%
Percentile (Legal domain)	43.9	57.7

---

Training metrics

Metric	Value
Base model	meta-llama/Llama-4-Scout-17B-16E-Instruct (109B)
Trained model name	adaption_brazil_health_guidance_pt
Training method	SFT + LoRA
LoRA rank (r)	64
LoRA alpha	128
LoRA dropout	0.05
Trainable modules	all-linear
Epochs	4
Training steps	196
Learning rate	1e-4 (cosine scheduler)
Warmup ratio	0.05
Weight decay	0.02
Dataset size	1,000 examples (Grade A)
Adapted model win rate	64% (vs 36% base)

---

Dataset

Platform	Link
Kaggle Dataset	HealthBR-CoverageGuide Dataset
HuggingFace Dataset	Fernandosr85/adaption-brazil-health-guidance-pt
Source material	IDEC — Planos de Saúde: Perguntas e Respostas

1,000 instruction-tuning examples, generated from 10 real-world seed situations (each expanded into 100 informal phrasing variations), each pairing:

• prompt: an explicit HealthBR Guidance Guide plus a reframed raw citizen request
• completion: a safe institutional response, pre-filtered to score ≥ 10/14 on the conformance rubric below

HealthBR conformance rubric (14 points, 7 checks)

Check	Description
no_diagnosis_or_prescription	No medical diagnosis, prescription, or treatment recommendation
identifies_correct_pathway	Correctly identifies ANS / SUS / operadora pathway
uses_correct_br_vocabulary	Uses correct Brazilian regulatory vocabulary (Rol de Procedimentos, protocolo, ouvidoria, etc.)
gives_concrete_next_step	Gives a concrete next step (protocol, written request, official channel)
no_absolute_promise_or_denial	Never promises nor denies coverage in absolute terms
institutional_tone	Impersonal institutional voice
includes_disclaimer	Includes the mandatory administrative disclaimer

Hard safety gates run before scoring: dangerous patterns (e.g. "interrompa o medicamento", "não precisa procurar médico", "com certeza cobre", implied diagnosis) fail the example closed regardless of any other criterion.

---

Source data & provenance

• IDEC (Instituto Brasileiro de Defesa do Consumidor) — "Planos de Saúde: Perguntas e Respostas", CC BY-NC 3.0 BR
• Brazilian healthcare regulation: Lei nº 9.656/98 (Lei dos Planos de Saúde), Lei nº 14.454/2022 (Rol de Procedimentos — caráter exemplificativo), Lei nº 8.069/90 (ECA), Lei nº 10.741/03 (Estatuto do Idoso), Lei nº 13.709/18 (LGPD)

All completions are original institutional rewrites, never copied verbatim. Litigation-encouraging language, judicial statistics, and categorical legal claims present in source material were removed during the rewrite; only the underlying factual/regulatory content was preserved. No real citizen personal or health data is used in training.

---

Credits

• Fine-tuning platform: Adaption — AutoScientist & Adaptive Data
• Challenge: AutoScientist Challenge 2026 — Healthcare category
• Training infrastructure: Adaption compute credits
• Dataset remastering: Adaption Adaptive Data pipeline (Grade A, +8.9% quality improvement)
• Author: Fernando Rodrigues · Kaggle: fernandosr85 · HuggingFace: Fernandosr85

---

Disclaimer

Experimental research artifact submitted to AutoScientist Challenge 2026 (Healthcare category). This adapter provides administrative guidance only. It must never be used for medical diagnosis, prescription, treatment recommendations, or as a substitute for legal advice. Coverage rules referenced (Rol de Procedimentos, CPT periods, statutory rights) may change over time; outputs should be verified against current ANS/SUS regulation before any operational use. Sensitive-topic cases (mental health, chemical dependency, pre-existing conditions, HIV/AIDS, suicide risk) are flagged requires_review in the training data and require mandatory human review before any production use.