Benchmark
Every number traces to a run id. Synthetic encounters use templated dialogue with exact answer keys; tune on dev, report test. Oracle extraction (hand-annotated mentions) is the engine's upper bound; LLM extraction is the real system.
Every number traces to a run id. Synthetic encounters use templated dialogue with exact answer keys; tune on dev, report test. Oracle extraction (hand-annotated mentions) is the engine's upper bound; LLM extraction is the real system.