The authors propose KnowledgeDebugger, a GUI-based exploration tool designed to support the initial phase of research into how Transformers store and process knowledge. This tool facilitates the exploration of knowledge localization and editing on individual samples before conducting more statistically robust experiments.
- It offers no-code access to methods in EasyEdit, a widely used library of state-of-the-art Knowledge Editing approaches.
- The tool is inspired by LM-Debugger and supports case studies of recent findings in the field.