Core ML Concepts · advanced · concept 48 of 311 · Updated Nov 4, 2026

🎯Hard Negative Mining

Hard negative mining selects, for each training example, the wrong answers the current model finds most similar to the right one, and trains against them. It matters for embedding, retrieval and face models because random negatives are too easy and stop producing useful gradient. DPR used top BM25 passages that lacked the answer; SimCSE uses contradiction sentences. The risk is false negatives: a mined negative that is actually relevant.