We use cookies to improve your experience on our site.
AVID-2025-R0004
Description
Evaluation of the LLM gpt-4o-mini-2024-07-18 on the wmdp_bio benchmark using Inspect Evals
Details
WMDP is a dataset of 3,668 multiple-choice questions surrounding hazardous knowledge in biosecurity, cybersecurity, and chemical security. WMDP serves as both a proxy evaluation for hazardous knowledge in large language models (LLMs) and a benchmark for unlearning methods to remove such knowledge.
We evaluated the LLM gpt-4o-mini-2024-07-18 on this benchmark.
Metrics
A simple accuracy is calculated over the datapoints.
| Scorer | Metric | Value |
|---|---|---|
| choice | accuracy | 0.77 |
| choice | stderr | 0.01 |
References
AVID Taxonomy Categorization
- Risk domains: Performance
- SEP subcategories: P0402: Physical Safety
- Lifecycle stages: L05: Evaluation
Affected or Relevant Artifacts
- Developer: OpenAI
- Deployer: OpenAI
- Artifact Details:
| Type | Name |
|---|---|
| Model | gpt-4o-mini-2024-07-18 |
Other information
- Report Type: Measurement
- Credits: Harsh Raj
- Date Reported: 2025-05-26
- Version: 0.2
- AVID Entry