## Held-out labels (mean): V-measure / many-to-one / labels / share in a class / sense / sense-COLL / merges (runs at the 2000 cap)
3000 LV-I: 0.606 / 0.676 / 897.900 / 0.633 / 0.722 / 0.744 / 271.1 (0)
3000 LV-A: 0.712 / 0.904 / 887.200 / 0.236 / 0.822 / 0.822 / 366.0 (0)
10000 LV-I: 0.615 / 0.682 / 938.400 / 0.625 / 0.652 / 0.663 / 356.5 (0)
10000 LV-A: 0.735 / 0.941 / 948.100 / 0.206 / 0.865 / 0.866 / 388.1 (0)
30000 LV-I: 0.570 / 0.601 / 807.900 / 0.734 / 0.530 / 0.543 / 627.8 (0)
30000 LV-A: 0.733 / 0.921 / 907.400 / 0.249 / 0.835 / 0.834 / 435.9 (0)
## Re-parse (mean): link F1 / ucomp type recall / ucomp precision; demo-setting LV-I F1; L-MST F1
3000 LV-I: 0.454 / 0.165 / 0.304 demo 0.449 L-MST 0.626
3000 LV-A: 0.467 / 0.131 / 0.313 demo 0.462 L-MST 0.626
10000 LV-I: 0.515 / 0.232 / 0.293 demo 0.504 L-MST 0.689
10000 LV-A: 0.534 / 0.202 / 0.301 demo 0.528 L-MST 0.689
30000 LV-I: 0.545 / 0.270 / 0.259 demo 0.544 L-MST 0.700
30000 LV-A: 0.589 / 0.273 / 0.302 demo 0.572 L-MST 0.700
## Predictions (at 30000)
C1: FAIL — share of held-out tokens in a class 0.734 (bar 0.8)
C2: FAIL — sense LV-I − LV-I-COLL -0.013 [-0.043, +0.016]
C3: FAIL — link F1 LV-I converged − demo setting +0.001 [-0.014, +0.017]
C4: FAIL — link F1 LV-I − L-MST -0.155 [-0.170, -0.140] (bar: lower > -0.02)