DataShield: Detecting Risky Fine-Tuning Data Across LLMs via Consensus Subspace Alignment

arXiv·medium signal

Fine-tuning on even benign task-specific data can silently erode an LLM's safety, and prior detectors rely on a single mean vector tied to one model and tokenizer. DataShield instead builds a consensus safety subspace aligned across multiple models, making risky-data detection transferable rather than model-specific. Practical for teams doing domain adaptation who need to screen datasets for safety degradation before shipping a fine-tune.