Skip to main content
. 2023 Apr 14;4(4):100729. doi: 10.1016/j.patter.2023.100729

Table 6.

Comparison of test performance (and standard deviation) on the BIOSSES and BioASQ tasks with major layer-specific adaptation methods, all with LARGE models

Baseline Layer reeze Layerwise decay Layer reinit
BIOSSES

BioBERT-LARGE 84.92 (10.2) 88.65 (6.60) 90.13 (1.70) 91.53 (4.09)
BlueBERT-LARGE 82.35 (2.21) 84.56 (1.95) 84.80 (2.58) 86.18 (1.21)
PubMedBERT-LARGE 91.06 (1.51) 91.19 (1.12) 90.87 (0.92) 92.73 (0.96)
PubMedELECTRA-LARGE 71.61 (4.91) 86.42 (3.33) 86.17 (1.21) 90.33 (1.04)

BioASQ

BioBERT-LARGE 67.79 (6.59) 74.57 (3.18) 78.93 (3.39) 74.43 (9.76)
BlueBERT-LARGE 70.43 (3.91) 70.21 (2.99) 72.21 (2.68) 70.86 (2.15)
PubMedBERT-LARGE 92.36 (1.36) 93.14 (2.35) 93.36 (1.22) 91.21 (1.54)
PubMedELECTRA-LARGE 79.93 (5.04) 88.64 (4.09) 93.14 (1.71) 88.07 (3.21)

Highest performance for model (row).