HyperCLOVA X Technical Report
Naver's technical report for its HyperCLOVA X large language model family. The report's ethics-principles section names 'self-anthropomorphism' — the model presenting a human persona, emotions, or relationships with humans that could cause a user to misunderstand it as a real human — as a prohibited harmful-content category, alongside child safety, alongside more conventional categories such as advice on criminal or dangerous behavior and sexual content.
Key Findings
- Ethics principles for content safety explicitly list 'self-anthropomorphism... such as human persona, emotions, and relationships with humans that can cause user to misunderstand them as real human' as a named harmful-content category
- Child safety is listed as a separate named harmful-content category within the same ethics-principles framework
- The self-anthropomorphism category sits alongside more conventional categories including advice on criminal/dangerous behavior, violence and cruelty, sexual content, and anti-ethical/normative content
Methodology Notes
General technical report covering model architecture, training, and evaluation across many capability and safety dimensions; the anthropomorphism/child-safety content is a small named-category section within a much broader document, not a dedicated companionship-safety study. Submitted to arXiv April 2, 2024; a revised version (v2) was posted April 13, 2024.
Sources
arXiv abstract (primary)
Archived snapshot (Wayback Machine) — preserved against link rot
Authors
Kang Min Yoo, Jaegeun Han, Sookyo In, Heewon Jeon, Jisu Jeong, Jaewook Kang, Hyunwook Kim, Kyung-Min Kim, Munhyong Kim
Tags
Cite This
APA
Kang Min Yoo et al. (2024). HyperCLOVA X Technical Report. Naver. https://arxiv.org/abs/2404.01954