## Token classes (mean): V-measure / many-to-one / sense accuracy
tier 3000
JT-V1 0.614 / 0.619 / 0.523
TYPE-A 0.512 / 0.567 / 0.482
TYPE 0.809 / 0.663 / 0.655
MAJ 0.975 / 0.987 / 0.875
JT-V1-SHUF 0.478 / 0.512 / 0.364
TYPE-A-SHUF 0.435 / 0.496 / 0.365
tier 10000
JT-V1 0.761 / 0.757 / 0.630
TYPE-A 0.664 / 0.701 / 0.577
TYPE 0.883 / 0.755 / 0.695
MAJ 0.977 / 0.988 / 0.885
JT-V1-SHUF 0.582 / 0.607 / 0.447
TYPE-A-SHUF 0.515 / 0.567 / 0.422
tier 30000
JT-V1 0.842 / 0.829 / 0.637
TYPE-A 0.795 / 0.804 / 0.593
TYPE 0.901 / 0.771 / 0.642
MAJ 0.977 / 0.988 / 0.886
JT-V1-SHUF 0.709 / 0.722 / 0.549
TYPE-A-SHUF 0.640 / 0.680 / 0.505
## Parse link F1 (mean)
3000: ATTN-V1 0.485 WORD-MI 0.158 GOLD-MI 0.699
10000: ATTN-V1 0.572 WORD-MI 0.272 GOLD-MI 0.734
30000: ATTN-V1 0.625 WORD-MI 0.378 GOLD-MI 0.734
## Continuous checkpoints, tier 30000 (mean V-measure)
1000:0.517 2000:0.580 4000:0.653 8000:0.746 16000:0.803 30000:0.842
## Predictions
J1: PASS — sense JT-V1 − TYPE-A at 30000: +0.044 [+0.030, +0.058]
J1+: FAIL — sense JT-V1 − MAJ at 30000: -0.248 [-0.368, -0.129]
J2: FAIL — V-measure JT-V1 0.842 (bar 0.7); − TYPE -0.059 [-0.127, +0.010]
J3: PASS — ATTN-V1 − WORD-MI 3000: +0.327 [+0.319]; 10000: +0.300 [+0.283]; 30000: +0.247 [+0.237]
J4': PASS — checkpoints from 4K: 0.653 0.746 0.803 0.842
J5: PASS — 60/60 audit checks PASS
J6: FAIL — shuffled sense JT-V1 − TYPE-A at 30000: +0.045 [+0.032, +0.057]