Loading...
標記為「subword」的 1 篇文章
利用ing LLM tokenization quirks to smuggle harmful content past safety filters by manipulating how text is split into tokens at the subword level.