LoRA-abliterated

transferpost-training

what

• a recipe for transfer uncensoring
• steps:
  1. Extraction: Extract a LoRA adapter by comparing two models: a censored model (e.g. Llama 3) and an abliterated model.
  2. Merge: Merge this new LoRA adapter using task arithmetic to a censored Llama 3.1 to abliterate it.