[2409.19663] Identifying Knowledge Editing Types in Large Language Models

View a PDF of the paper titled Figuring out Data Enhancing Varieties in Massive Language Fashions, by Xiaopeng Li and seven different authors

View PDF
HTML (experimental)

Summary:Data enhancing has emerged as an environment friendly expertise for updating the data of enormous language fashions (LLMs), attracting growing consideration in recent times. Nonetheless, there’s a lack of efficient measures to stop the malicious misuse of this expertise, which may result in dangerous edits in LLMs. These malicious modifications may trigger LLMs to generate poisonous content material, deceptive customers into inappropriate actions. In entrance of this threat, we introduce a brand new process, Data Enhancing Sort Identification (KETI), geared toward figuring out various kinds of edits in LLMs, thereby offering well timed alerts to customers when encountering illicit edits. As a part of this process, we suggest KETIBench, which incorporates 5 varieties of dangerous edits protecting hottest poisonous sorts, in addition to one benign factual edit. We develop 4 classical classification fashions and three BERT-based fashions as baseline identifiers for each open-source and closed-source LLMs. Our experimental outcomes, throughout 42 trials involving two fashions and three data enhancing strategies, reveal that every one seven baseline identifiers obtain respectable identification efficiency, highlighting the feasibility of figuring out malicious edits in LLMs. Extra analyses reveal that the efficiency of the identifiers is impartial of the reliability of the data enhancing strategies and reveals cross-domain generalization, enabling the identification of edits from unknown sources. All information and code can be found in this https URL. Warning: This paper incorporates examples of poisonous textual content.