BABEL Codec: Fully Opens the Interior of the GPT-2 Model
The BABEL codec translates the internal state of the GPT-2 model into English, achieving 94.7% behavior reconstruction.
The BABEL codec is the first complete and certified decoding tool that fully reveals all processes inside the GPT-2 small model. It can read the model's internal state into English and write English back into the model. The achieved 94.7% behavior reconstruction holds across every layer depth and text regime tested.