Results

Effectiveness

We apply GlitchHunter to eight open-source LLMs: GPT2-small, GPT2-xl, Llama2-7b-chat, Llama2-13b-chat, ChatGLM-6b, ChatGLM2-6b, Mistral-7b-Instruct, Vicuna-13b. The result against several baselines is shown as follows.

Efficiency

We compare the running time and token consumption of GlitchHunter against traversing the whole token set of the model. The results on all eight models are shown below:

Page updated

Google Sites

Report abuse