Table 8.
Ablation study on the impact of model pruning by comparing test performance on BLURB tasks after removing top layers of PubMedBERT
| Layers removed |
Performance drop | ||||
|---|---|---|---|---|---|
| 0 | 2 | 4 | 6 | ||
| BC5-chem | 93.33 | 93.22 | 92.96 | 92.40 | −0.93 |
| BC5-disease | 85.62 | 85.35 | 85.17 | 84.29 | −1.33 |
| NCBI-disease | 87.82 | 88.38 | 87.99 | 87.38 | −0.44 |
| BC2GM | 84.52 | 84.32 | 83.46 | 82.05 | −2.47 |
| JNLPBA | 79.10 | 78.94 | 78.88 | 78.07 | −1.03 |
| EBM PICO | 73.38 | 73.38 | 73.33 | 73.35 | −0.05 |
| ChemProt | 77.24 | 76.11 | 73.40 | 72.74 | −4.50 |
| DDI | 82.36 | 82.16 | 79.71 | 79.30 | −3.06 |
| GAD | 83.96 | 82.33 | 80.23 | 79.21 | −4.75 |
| BIOSSES | 92.30 | 92.66 | 92.80 | 92.12 | −0.18 |
| HoC | 82.32 | 82.46 | 82.43 | 82.01 | −0.31 |
| PubMedQA | 55.84 | 51.22 | 49.76 | 50.08 | −6.08 |
| BioASQ | 87.56 | 83.73 | 77.50 | 74.00 | −13.56 |