PKU Alignment Group / Center for AI Safety and Governance, Peking University
- Type
- Academic institution
- Lieu
- Chine — China (Beijing)
- Dernière vérification
- 2026-09-22
- Prochaine vérification
- 2027-03-21
Pas encore traduit — affiché en anglais.
Comment les joindre
- PKU Alignment Group general contact
- GitHub — open-source alignment datasets and frameworks, open to contribution
- Recruitment form
- PKU Center for AI Safety and Governance
- Homepage
Ce qu'il fait
PKU-Alignment is a student-organised research team at Peking University working on reinforcement learning, LLMs, world models and AI alignment. Widely known for open-sourcing alignment resources — PKU-SafeRLHF datasets, Safe-RLHF training code, BeaverTails and the Align-Anything framework — published at ACL, NeurIPS and ICLR, with collaborations including Anthropic, Google DeepMind, DeepSeek and Alibaba. PKU's Institute for AI separately hosts a Center for AI Safety and Governance.
Évaluation franche
Notably open for a Chinese institution — a public Gmail address, open GitHub, and a public recruitment form. Its open-source alignment work is genuinely used worldwide, making GitHub the highest-bandwidth channel. Student-run, so responsiveness varies with cohort.
Préoccupations couvertes
Value alignment of LLMs — safe RLHF, helpfulness/harmlessness tradeoffs, and increasingly scalable oversight. Genuinely alignment-focused, not a near-term-harms group.
Canal gouvernemental
Indirect — PKU researchers feed into Chinese AI safety standards discussions and the China AI Safety and Development Network.