Pretraining data curation targets safer alignment

This is a real research direction, but it looks more like safety shaping than concept deletion. Only very separable concepts are plausibly ablatable; for most semantic targets there is no demonstrated zero floor, just partial reductions and leakage through benign context. RealToxicityPrompts and newer safety-pretraining work show data-centric pretraining can materially reduce toxic outputs, but not eliminate them. A Pretrainer's Guide to Training Data found a real safety-versus-capability tradeoff as toxicity filters get stricter. Personas as a Way to Model Truthfulness in Language Models suggests truthfulness is learned from data structure, so isolated deletions will not fully erase deception. The biggest risk is entanglement with legitimate scientific, code, and historical text, which can quietly dent scientific and algorithmic capability if curation is too blunt.