Why not train on all two million synthetic query-passage pairs generated by an LLM?

Training on unfiltered synthetic data introduces noise. A round-trip check—training a preliminary retriever on the set and keeping only pairs where the retriever recovers the source—acts as an effective filter. This process increases precision (from π to πα/r) without the need for human labels or the overhead and potential blind spots of using an LLM judge to score each pair.