{ "seed": 2951, "math_check": { "max_role_contraction_abs_error": 1.3322676295501878e-15, "exact_query_top1": true, "exact_query_probability": 0.9999999999656395, "exact_rebinding_l2_error": 1.161975694124784e-10, "exact_entropy": 8.795314964741619e-10, "mismatched_role_entropy": 0.0008730161579515994 }, "benchmark": { "structured_acc": 0.995, "dense_acc": 0.579, "structured_err": 0.03480833023786545, "dense_err": 2.383237838745117, "structured_entropy": 0.005550319146098912, "dense_entropy": 0.06881254905095009, "wrong_entropy": 0.20854220871185467 }, "benchmark_repeat": { "structured_acc": 0.995, "dense_acc": 0.551, "structured_err": 0.043424125760793686, "dense_err": 2.5534236431121826, "structured_entropy": 0.009825814736692964, "dense_entropy": 0.07332719177381702, "wrong_entropy": 0.180572374217896 } }