Re:Deepseek does not want to talk about Xi Jinping
by Salgat on Tuesday January 28, 2025 @01:32AM
(#42849489)
I more meant, if a model is only trained on 100% absolutely positive data concerning a specific subject, including any data that implies certain taboos it must avoid at all costs, the model will embed that strong bias in its learned parameters.
More-likely it is some combination of a system prompt and output-filtering code.