You removed the refusals. Did you check what else came with the model? Safety matters just as much when you run abliterated AI. Choosing fewer refusals still means taking responsibility for the files, templates and code you bring onto your machine. This is why @ExileAI_0's model audit deserves attention. It inspects GGUF and safetensors artifacts for known suspicious chat-template patterns, structural inconsistencies and checksum mismatches. It can flag scripts shipped alongside a model, check hashes against a pinned Hugging Face revision, and create per-tensor fingerprints for comparison with a trusted baseline. The example that stuck with me: replace a chat template, leave the weights unchanged, and the audit raises a critical finding. The instructions surrounding a model deserve scrutiny too. What makes this repo especially valuable is its honesty: matching hashes establish consistency with a reference. They cannot prove the weights are benign. The reports explain what was checked and what remains unknown, in plain language. That is the kind of practical security work the local AI community needs more of. More freedom should come with better verification. Big credit to @ExileAI_0 for building this. If you download, quantize or share models, give it a look. github.com/ExileAI/model-aud…