@grimjim on Hugging Face: "In principle, it's possible to "abliterate" refusals in any Llama 3.1 8B…"

Hugging Face

Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Back to feed

grimjim

posted an update Jul 25

Post

2285

In principle, it's possible to "abliterate" refusals in any Llama 3.1 8B models via application of a LoRA, using only mergekit.

Proof of concept below:
grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter

grimjim

Jul 25

Select GGUF quants.
https://huggingface.co/grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter-GGUF

ZeroWw

Jul 26

would be nice to see how the lora was created, or providing a colab like mlabonne did.

grimjim

Jul 26

An example use of mergekit to apply the lora is documented at the bottom of the model card as well as in mergekit_config.yaml. Although task_arithmetic was used, a passthrough merge will work as well.

An example of the mergekit LoRA extraction command is at the bottom of the model card for the LoRA:
https://huggingface.co/grimjim/Llama-3-Instruct-abliteration-LoRA-8B

In this post

grimjim Jim Lai
ZeroWw Robert Sinclair