Code & Datasets

Code

Code of GlitchHunter is available here.

Ground Truth of Glitch Tokens

We present the ground truth of glitch tokens in seven LLMs collected in the empirical study here.

Embeddings of Open-source LLMs

We present eight embeddings of open-source LLMs: GPT2-small, GPT2-xl, Llama2-7b-chat, Llama2-13b-chat, ChatGLM-6b, ChatGLM2-6b, Mistral-7b-Instruct, Vicuna-13b. The data and the usage of .h5ad file is described here.