The Cyberspace Administration of China: When providers carry out activities such as pre-training and optimization training, they should strengthen the management of training data.

The Cyberspace Administration of China has solicited public opinions on the Interim Measures for the Management of Artificial Intelligence Human-like Interactive Services. It mentions that when providers carry out activities such as pre-training and optimization training, they should strengthen the management of training data and comply with the following regulations: use datasets that conform to the core socialist values and embody the excellent traditional Chinese culture; clean and label the training data to enhance transparency and reliability, prevent data poisoning, data tampering, and other behaviors; increase the diversity of training data, enhance the security of model-generated content through negative sampling, adversarial training, and other means; evaluate the security of synthetic data when used for model training and key capability optimization; strengthen daily inspection of training data, regularly iterate and upgrade the data, continuously optimize the performance of products and services; ensure the legal and traceable sources of training data, take necessary measures to ensure data security, and prevent data leakage risks.