SafetyBench is a benchmark dataset with 11,435 multiple-choice questions designed to evaluate the safety of large language models. The dataset spans 7 distinct categories of safety concerns and incorporates both Chinese and English data. It was created by thu-coai and last updated on September 14, -2023.
Use Cases
- Benchmarking LLM safety performance based on 7 distinct categories of safety concerns.
- Evaluating multilingual model safety using the incorporated Chinese and English data.
- Analyzing model failure modes across different safety question types.
Strengths
- Contains 11,435 diverse multiple-choice questions.
- Covers 7 distinct categories of safety concerns.
- Includes data in both Chinese and English languages.
Limitations
- Row count for individual test splits is unknown.
- Column-level documentation is absent; field semantics must be inferred after download.
- Last updated 2023-09-14 05:25:39; freshness should be verified.
Provenance
- Source
- thu-coai
- Collection Method
- Likely curated for research purposes as a benchmark.
- Time Range
- null
- Freshness
- Last updated 2023-09-14 05:25:39.
- Geography
- null