li-lab/First-do-NOHARM-v2

Japanese First Do NOHARM v2

This dataset is a human-reviewed Japanese adaptation of the First Do NOHARM v2 benchmark for evaluating the safety of large language models in clinical decision-making.

Dataset Overview

The dataset contains 330 Japanese clinical prompts, consisting of:

• 30 baseline cases
• 300 perturbation items (10 perturbations per baseline case)

Each baseline case and its 10 perturbations share the same rubric and matching guidance. Perturbations modify specific aspects of the clinical scenario while keeping the underlying case and evaluation criteria consistent.

The Japanese prompts were human-reviewed to ensure that the clinical meaning and intended perturbations were preserved.

Dataset Structure

Each row corresponds to one baseline case or one perturbation.

Column	Description
id	Unique identifier for each item
base_id	Identifier of the base case shared by the baseline item and its perturbations
perturb_id	Perturbation identifier; null for baseline items
category	Item category, such as baseline or perturbation
candidate_id	Candidate identifier, when applicable
open	Whether the item belongs to the open subset
prompt	Human-reviewed Japanese clinical prompt presented to the model
rubric	Full evaluation rubric for the base case; populated only in the baseline row
guidance	Matching guidance for the base case; populated only in the baseline row

Baseline and Perturbation Relationship

Each base case has the following structure:

base case
├── baseline item
├── perturbation 1
├── perturbation 2
├── ...
└── perturbation 10