The contrastive corpora at the start of every abliteration pipeline. Difference-of-means over these pairs produces the direction that later gets orthogonalized out of the model weights. Small sets (hundreds of prompts per side) are the community standard.
git clone https://github.com/Sumandora/remove-refusals-with-transformers cd remove-refusals-with-transformers pip install -r requirements.txt python compute_refusal_dir.py --model Qwen/Qwen3-4B-Instruct-2507 --samples 512
Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.