byroneverson/glm-4-9b-chat-abliterated

glm-4-9b-chat-abliterated

Version 1.1 (Updated 9/1/2024): Layer 16 is used for abliteration instead of 20. Refusal mitigation tends to work better with this layer. PCA and cosine similarity tests seem to agree.

Check out the <a href="https://huggingface.co/byroneverson/glm-4-9b-chat-abliterated/blob/main/abliterate-glm-4-9b-chat.ipynb">jupyter notebook</a> for details of how this model was abliterated from glm-4-9b-chat.

The python package "tiktoken" is required to quantize the model into gguf format. So I had to create <a href="https://huggingface.co/spaces/byroneverson/gguf-my-repo-plus-tiktoken">a fork of GGUF My Repo (+tiktoken)</a>.