SciSafeEval: A Comprehensive Benchmark for Safety Alignment of Large Language Models in Scientific Tasks

SciSafeEval is a comprehensive benchmark designed to assess the safety alignment of Large Language Models (LLMs) in scientific tasks. The benchmark comprises 31,840 harmful queries spanning four major scientific domains: chemistry, biology, medicine, and physics, encompass textual, molecular, protein, and genomic languages.

• 📑 Paper: https://arxiv.org/pdf/2410.03769
• 👨‍💻 Code: https://github.com/DavidLee528/SciSafeEval
• 🌐 Website: https://scisafeeval.tianhao.li

Dataset Configurations

This dataset is organized into multiple configurations due to different schema structures across scientific domains:

• biology: Contains protein and gene-related tasks with the sequence field
• chemistry: Contains molecule-related tasks with smiles and selfies fields
• medicine: Contains medical inference and knowledge retrieval tasks
• physics: Contains physics knowledge retrieval tasks

To load a specific configuration:

from datasets import load_dataset
# Load the biology configuration
dataset = load_dataset("Tianhao0x01/SciSafeEval", "biology")
# Load the chemistry configuration
dataset = load_dataset("Tianhao0x01/SciSafeEval", "chemistry")

Sample Data

Construction Method

Experiment Result