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 jupyter notebook 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 fork of GGUF My Repo (+tiktoken).
