Datasets that mention abliteration or refusal-adjacent concepts but do not fit the five workflow stages. Includes concept-erasure corpora, steering-vector datasets, quantization-repair imatrix sets, and other single-purpose material.
git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp && make ./llama-imatrix -m model.gguf -f imatrix_calib.txt -o imatrix.dat ./llama-quantize --imatrix imatrix.dat model.gguf model-IQ4_NL.gguf IQ4_NL
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.