Definition
SNR-aware filtering selects training data based on signal-to-noise ratio — keeping examples where the learning signal is high relative to label noise or task ambiguity. It’s a common trick in RL and preference training, where a small amount of clean signal beats a large amount of muddy data.