We use cookies to improve your experience on our site.
AVID-2022-R0005
Description
Demographic bias found in EleutherAI/gpt-neo-125M for multiple sensitive categories, as measured on prompts supplied in the BOLD dataset
Details
When supplied with prompts from the BOLD dataset, toxicity scores for completed sentences tagged with the ‘Islam’ religion demonstrate a higher degree of toxicity than other religions.
References
- Gender Bias Evaluation for Masked Language modelling: BOLD
- EleutherAI/gpt-neo-125M on Hugging Face
- BOLD: Dataset and Metrics for Measuring Biases in Open-Ended Language Generation
AVID Taxonomy Categorization
- Risk domains: Ethics
- SEP subcategories: E0101: Group Fairness
- Lifecycle stages: L05: Evaluation
Affected or Relevant Artifacts
- Developer: EleutherAI
- Deployer: HuggingFace
- Artifact Details:
Type Name Model EleutherAI/gpt-neo-125M
Other information
- Report Type: Detection
- Credits: Subho Majumdar, AVID
- Date Reported: 2022-11-09
- Version: 0.1
- AVID Entry