Datasets:

Tropic-AI
/

CondutaBench

language:
  - pt
license: cc-by-4.0
task_categories:
  - question-answering
  - text-generation
pretty_name: CondutaBench
size_categories:
  - n<1K
tags:
  - medical
  - clinical-guidelines
  - evaluation
  - benchmark
  - brazilian-portuguese
configs:
  - config_name: questions
    default: true
    data_files:
      - split: test
        path: data/questions.jsonl
  - config_name: rubrics
    data_files:
      - split: test
        path: data/rubrics.jsonl
  - config_name: support
    data_files:
      - split: test
        path: data/support.jsonl

CondutaBench

715 Portuguese clinical-management questions, 4,722 rubric criteria, 14 clinical domains.

Code, evaluation scripts and complete instructions

CondutaBench evaluates clinical-management answers against Brazilian clinical guidelines. Clinical scenarios and questions are in Portuguese. Rubric labels are the final labels used in the paper; selected supporting passages were consolidated on October 13, 2026.

Load the data

from datasets import load_dataset

questions = load_dataset("Tropic-AI/CondutaBench", "questions", split="test")
rubrics = load_dataset("Tropic-AI/CondutaBench", "rubrics", split="test")
support = load_dataset("Tropic-AI/CondutaBench", "support", split="test")

Each configuration has 715 rows. Join records by id; each ID identifies the same question across configurations. The entire dataset is for evaluation, with no training partition.

Configuration	Contents
`questions`	`id`, `guideline_id`, `topic`, `scenario`, `question`
`rubrics`	`id`, `reference_answer`, ordered `criteria` (text, type, impact, weight, original reference)
`support`	`id`, `guideline_id`, selected `passages`, known support `limitations`

Evaluation

Give evaluated models only the scenario and question for closed-book evaluation. Do not include the reference answer, rubrics or criterion-specific support passages in model inputs. Source documents for full-context and agentic evaluation must be obtained separately: complete guideline documents are not distributed here.

The GitHub repository provides answer generation and evaluation through Maritaca, OpenAI, Anthropic and OpenRouter, plus offline scoring. The default judge is Sabiá-4-Thinking; other judges are configurable. API requests may incur charges under your own account.

Score is the sum of weights of satisfied/triggered criteria divided by the sum of positive weights, multiplied by 100. Scores are not clipped. Correct requires every essential criterion and no negative criterion; Complete requires every positive criterion and no negative criterion. A triggered negative criterion means the answer committed the corresponding error or prohibited action. Aggregate metrics average over questions.

Data construction and limitations

Questions and rubrics were generated from clinical guidelines and reviewed by clinicians. Criterion types and impact labels underwent further review combining clinical annotation and model agreement. Labels are fixed; the support file stores later-selected passages without changing criterion wording or weights. See the repository for the schema and evaluation instructions.

Support passages can be partial. limitations records known gaps; no recorded limitation does not guarantee complete source support. The evidence file is not an independent set of reference answers. Automatic judge scores can contain errors and depend on the judge used.

The benchmark uses guideline-grounded clinical scenarios for research evaluation. It is not a clinical decision tool. Its results do not establish safety or effectiveness in patient care.

License

Original benchmark content is released under CC BY 4.0. Selected third-party guideline excerpts retain their original authorship and rights and are not relicensed by this release. The code in the linked GitHub repository is MIT licensed.

Citation

The paper will be publicly available soon. Citation details will be updated upon publication.