workmem/validation/validation_step_500.jsonl

{"global_idx": 0, "total_positions": 437, "gt_mem_token_count": 11, "pred_mem_token_count": 223, "correct_mem_token_count": 3, "mem_token_precision": 0.013452914798206279, "mem_token_recall": 0.2727272727272727, "mem_token_accuracy": 0.2727272727272727, "slot_norm_mean": 198.0, "slot_sim_mean": 1.0, "rank": 0}
{"global_idx": 1, "total_positions": 594, "gt_mem_token_count": 5, "pred_mem_token_count": 277, "correct_mem_token_count": 1, "mem_token_precision": 0.0036101083032490976, "mem_token_recall": 0.2, "mem_token_accuracy": 0.2, "slot_norm_mean": 199.0, "slot_sim_mean": 0.99609375, "rank": 0}
{"global_idx": 2, "total_positions": 638, "gt_mem_token_count": 5, "pred_mem_token_count": 456, "correct_mem_token_count": 4, "mem_token_precision": 0.008771929824561403, "mem_token_recall": 0.8, "mem_token_accuracy": 0.8, "slot_norm_mean": 196.0, "slot_sim_mean": 0.99609375, "rank": 0}
{"global_idx": 3, "total_positions": 408, "gt_mem_token_count": 6, "pred_mem_token_count": 179, "correct_mem_token_count": 3, "mem_token_precision": 0.01675977653631285, "mem_token_recall": 0.5, "mem_token_accuracy": 0.5, "slot_norm_mean": 193.0, "slot_sim_mean": 1.0, "rank": 0}
{"global_idx": 4, "total_positions": 447, "gt_mem_token_count": 8, "pred_mem_token_count": 209, "correct_mem_token_count": 2, "mem_token_precision": 0.009569377990430622, "mem_token_recall": 0.25, "mem_token_accuracy": 0.25, "slot_norm_mean": 198.0, "slot_sim_mean": 1.0, "rank": 0}
{"global_idx": 5, "total_positions": 414, "gt_mem_token_count": 6, "pred_mem_token_count": 185, "correct_mem_token_count": 2, "mem_token_precision": 0.010810810810810811, "mem_token_recall": 0.3333333333333333, "mem_token_accuracy": 0.3333333333333333, "slot_norm_mean": 194.0, "slot_sim_mean": 1.0, "rank": 0}
{"global_idx": 6, "total_positions": 577, "gt_mem_token_count": 11, "pred_mem_token_count": 370, "correct_mem_token_count": 1, "mem_token_precision": 0.002702702702702703, "mem_token_recall": 0.09090909090909091, "mem_token_accuracy": 0.09090909090909091, "slot_norm_mean": 197.0, "slot_sim_mean": 0.9921875, "rank": 0}
{"global_idx": 7, "total_positions": 631, "gt_mem_token_count": 17, "pred_mem_token_count": 464, "correct_mem_token_count": 3, "mem_token_precision": 0.00646551724137931, "mem_token_recall": 0.17647058823529413, "mem_token_accuracy": 0.17647058823529413, "slot_norm_mean": 200.0, "slot_sim_mean": 0.9921875, "rank": 1}
{"global_idx": 8, "total_positions": 591, "gt_mem_token_count": 7, "pred_mem_token_count": 320, "correct_mem_token_count": 2, "mem_token_precision": 0.00625, "mem_token_recall": 0.2857142857142857, "mem_token_accuracy": 0.2857142857142857, "slot_norm_mean": 200.0, "slot_sim_mean": 0.99609375, "rank": 1}
{"global_idx": 9, "total_positions": 579, "gt_mem_token_count": 13, "pred_mem_token_count": 405, "correct_mem_token_count": 4, "mem_token_precision": 0.009876543209876543, "mem_token_recall": 0.3076923076923077, "mem_token_accuracy": 0.3076923076923077, "slot_norm_mean": 199.0, "slot_sim_mean": 1.0, "rank": 1}
{"global_idx": 10, "total_positions": 596, "gt_mem_token_count": 13, "pred_mem_token_count": 262, "correct_mem_token_count": 1, "mem_token_precision": 0.003816793893129771, "mem_token_recall": 0.07692307692307693, "mem_token_accuracy": 0.07692307692307693, "slot_norm_mean": 200.0, "slot_sim_mean": 0.99609375, "rank": 1}
{"global_idx": 11, "total_positions": 353, "gt_mem_token_count": 7, "pred_mem_token_count": 158, "correct_mem_token_count": 1, "mem_token_precision": 0.006329113924050633, "mem_token_recall": 0.14285714285714285, "mem_token_accuracy": 0.14285714285714285, "slot_norm_mean": 197.0, "slot_sim_mean": 1.0, "rank": 1}
{"global_idx": 12, "total_positions": 378, "gt_mem_token_count": 15, "pred_mem_token_count": 189, "correct_mem_token_count": 3, "mem_token_precision": 0.015873015873015872, "mem_token_recall": 0.2, "mem_token_accuracy": 0.2, "slot_norm_mean": 196.0, "slot_sim_mean": 0.99609375, "rank": 1}
{"global_idx": 13, "total_positions": 412, "gt_mem_token_count": 11, "pred_mem_token_count": 228, "correct_mem_token_count": 2, "mem_token_precision": 0.008771929824561403, "mem_token_recall": 0.18181818181818182, "mem_token_accuracy": 0.18181818181818182, "slot_norm_mean": 197.0, "slot_sim_mean": 1.0, "rank": 1}
{"global_idx": 14, "total_positions": 665, "gt_mem_token_count": 10, "pred_mem_token_count": 521, "correct_mem_token_count": 4, "mem_token_precision": 0.007677543186180422, "mem_token_recall": 0.4, "mem_token_accuracy": 0.4, "slot_norm_mean": 198.0, "slot_sim_mean": 1.0, "rank": 2}
{"global_idx": 15, "total_positions": 529, "gt_mem_token_count": 12, "pred_mem_token_count": 371, "correct_mem_token_count": 2, "mem_token_precision": 0.005390835579514825, "mem_token_recall": 0.16666666666666666, "mem_token_accuracy": 0.16666666666666666, "slot_norm_mean": 199.0, "slot_sim_mean": 0.9921875, "rank": 2}
{"global_idx": 16, "total_positions": 383, "gt_mem_token_count": 6, "pred_mem_token_count": 155, "correct_mem_token_count": 2, "mem_token_precision": 0.012903225806451613, "mem_token_recall": 0.3333333333333333, "mem_token_accuracy": 0.3333333333333333, "slot_norm_mean": 200.0, "slot_sim_mean": 0.99609375, "rank": 2}
{"global_idx": 17, "total_positions": 370, "gt_mem_token_count": 10, "pred_mem_token_count": 183, "correct_mem_token_count": 2, "mem_token_precision": 0.01092896174863388, "mem_token_recall": 0.2, "mem_token_accuracy": 0.2, "slot_norm_mean": 197.0, "slot_sim_mean": 0.99609375, "rank": 2}
{"global_idx": 18, "total_positions": 647, "gt_mem_token_count": 9, "pred_mem_token_count": 435, "correct_mem_token_count": 2, "mem_token_precision": 0.004597701149425287, "mem_token_recall": 0.2222222222222222, "mem_token_accuracy": 0.2222222222222222, "slot_norm_mean": 195.0, "slot_sim_mean": 0.9921875, "rank": 2}
{"global_idx": 19, "total_positions": 261, "gt_mem_token_count": 6, "pred_mem_token_count": 115, "correct_mem_token_count": 3, "mem_token_precision": 0.02608695652173913, "mem_token_recall": 0.5, "mem_token_accuracy": 0.5, "slot_norm_mean": 194.0, "slot_sim_mean": 0.99609375, "rank": 2}
{"global_idx": 20, "total_positions": 437, "gt_mem_token_count": 11, "pred_mem_token_count": 223, "correct_mem_token_count": 3, "mem_token_precision": 0.013452914798206279, "mem_token_recall": 0.2727272727272727, "mem_token_accuracy": 0.2727272727272727, "slot_norm_mean": 198.0, "slot_sim_mean": 1.0, "rank": 2}
{"global_idx": 21, "total_positions": 594, "gt_mem_token_count": 5, "pred_mem_token_count": 277, "correct_mem_token_count": 1, "mem_token_precision": 0.0036101083032490976, "mem_token_recall": 0.2, "mem_token_accuracy": 0.2, "slot_norm_mean": 199.0, "slot_sim_mean": 0.99609375, "rank": 2}
{"global_idx": 22, "total_positions": 302, "gt_mem_token_count": 6, "pred_mem_token_count": 119, "correct_mem_token_count": 2, "mem_token_precision": 0.01680672268907563, "mem_token_recall": 0.3333333333333333, "mem_token_accuracy": 0.3333333333333333, "slot_norm_mean": 194.0, "slot_sim_mean": 0.99609375, "rank": 3}
{"global_idx": 23, "total_positions": 366, "gt_mem_token_count": 6, "pred_mem_token_count": 153, "correct_mem_token_count": 1, "mem_token_precision": 0.006535947712418301, "mem_token_recall": 0.16666666666666666, "mem_token_accuracy": 0.16666666666666666, "slot_norm_mean": 198.0, "slot_sim_mean": 1.0, "rank": 3}
{"global_idx": 24, "total_positions": 662, "gt_mem_token_count": 14, "pred_mem_token_count": 392, "correct_mem_token_count": 2, "mem_token_precision": 0.00510204081632653, "mem_token_recall": 0.14285714285714285, "mem_token_accuracy": 0.14285714285714285, "slot_norm_mean": 198.0, "slot_sim_mean": 1.0, "rank": 3}
{"global_idx": 25, "total_positions": 362, "gt_mem_token_count": 8, "pred_mem_token_count": 224, "correct_mem_token_count": 2, "mem_token_precision": 0.008928571428571428, "mem_token_recall": 0.25, "mem_token_accuracy": 0.25, "slot_norm_mean": 199.0, "slot_sim_mean": 1.0, "rank": 3}
{"global_idx": 26, "total_positions": 466, "gt_mem_token_count": 13, "pred_mem_token_count": 194, "correct_mem_token_count": 1, "mem_token_precision": 0.005154639175257732, "mem_token_recall": 0.07692307692307693, "mem_token_accuracy": 0.07692307692307693, "slot_norm_mean": 199.0, "slot_sim_mean": 0.99609375, "rank": 3}
{"global_idx": 27, "total_positions": 712, "gt_mem_token_count": 7, "pred_mem_token_count": 461, "correct_mem_token_count": 3, "mem_token_precision": 0.006507592190889371, "mem_token_recall": 0.42857142857142855, "mem_token_accuracy": 0.42857142857142855, "slot_norm_mean": 197.0, "slot_sim_mean": 0.9921875, "rank": 3}
{"global_idx": 28, "total_positions": 631, "gt_mem_token_count": 17, "pred_mem_token_count": 464, "correct_mem_token_count": 3, "mem_token_precision": 0.00646551724137931, "mem_token_recall": 0.17647058823529413, "mem_token_accuracy": 0.17647058823529413, "slot_norm_mean": 200.0, "slot_sim_mean": 0.9921875, "rank": 3}
{"global_idx": 29, "total_positions": 591, "gt_mem_token_count": 7, "pred_mem_token_count": 320, "correct_mem_token_count": 2, "mem_token_precision": 0.00625, "mem_token_recall": 0.2857142857142857, "mem_token_accuracy": 0.2857142857142857, "slot_norm_mean": 200.0, "slot_sim_mean": 0.99609375, "rank": 3}