01:38:07 === error detector, SEQ=256 vocab=8192 === 01:40:40 lm 2000 loss=4.7330 (152s) 01:43:06 lm 4000 loss=3.6076 (298s) 01:45:26 lm 6000 loss=3.4091 (438s) 01:46:09 det 500 bce=0.4624 base=0.7838 (37s) 01:46:38 det 1000 bce=0.4412 base=0.7205 (66s) 01:47:08 det 1500 bce=0.4389 base=0.7674 (97s) 01:47:40 det 2000 bce=0.4396 base=0.7548 (128s) 01:47:51 ranking test 01:47:51 wrong fraction 0.7628 on 19760 positions 01:47:51 AUC detector=0.8220 1-confidence=0.9620 01:47:51 decode, mask rate 0.5 (draft K=8, 4 refill rounds, frac=0.20) 01:47:51 conf tok=0.5250 pair=0.3082 01:47:51 detector tok=0.5230 pair=0.3061 01:47:52 random tok=0.5201 pair=0.3065 01:47:52 decode, mask rate 0.9 (draft K=8, 4 refill rounds, frac=0.20) 01:47:52 conf tok=0.1537 pair=0.0674 01:47:52 detector tok=0.1543 pair=0.0666 01:47:52 random tok=0.1522 pair=0.0666 01:47:53 published pilots/detect.json