HealthBench-BR

A true/false benchmark of 1,780 paired clinical assertions grounded in Brazil's official clinical guidelines (PCDTs), published by the Ministry of Health.

For each clinical fact, two statements are provided: a true version faithful to the guideline, and a false version that modifies a single critical detail (a dosage, route of administration, monitoring interval, etc.). The dataset is perfectly balanced (50% true / 50% false), so correct classification requires precise factual recall — not surface heuristics. Evaluation extracts a Verdadeiro / Falso verdict from the model's response via regex.

• 1,780 items total
• 10 paired true/false statements per guideline · 178 guidelines total
• Train/test split by guideline (890/890 per split, 1780 total per split)

Usage

from datasets import load_dataset
ds = load_dataset("hugo/healthbench-br")

Results — open models from this project

Test-split accuracy (greedy decoding). All checkpoints are publicly available:

Model	Accuracy
Qwen2.5-14B-Instruct (baseline)	59.4
CPT + Replay (FineWeb-EN, 1:1)	63.8
RL only (LoRA, no CPT)	64.2
CPT (1 generator), LoRA	67.2
CPT (1 generator), full FT	69.6
CPT (4 generators), LoRA	70.3
CPT (4 generators), full FT	71.1
CPT (1 generator) + RL, LoRA	75.6
CPT (4 generators) + RL, LoRA ⭐	83.9

Frontier models (for reference)

Model	Accuracy
GPT-4.1	73.3
Google AI Overview	70.5
Claude Sonnet 4.6	77.6
GPT-5.2 (high)	78.5
Gemini 3.1 Pro	79.3

Citation

See the paper and code at the project repository:

Code & paper: https://github.com/hugoabonizio/clinical-protocols-br