SynthSentry for Detecting Synthetic Data Contamination in Training Corpora

SynthSentry provides a model-agnostic signal to detect synthetic data contamination by measuring lexical diversity collapse, n-gram tail truncation, and perplexity variance. This allows for screening unknown corpora before training to prevent model collapse caused by recursive training on synthetic outputs.