Continual Learning on MedCL 10 biomedical tasks
62.6OP ScoreO-LoRA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| O-LoRAType=architectural, Backbone=Gemma-2 9B-it, Epochs per task=3, Evaluation protocol=log-probability evaluation2026.06 | 62.6 | 61.3 | 0.066 | |
| ReplayReplay buffer size=10%, Backbone=Gemma-2 9B-it, Epochs per task=3, Evaluation protocol=log-probability evaluation2026.06 | 56.6 | 55.2 | 0.011 | |
| ReplayReplay buffer size=5%, Backbone=Gemma-2 9B-it, Epochs per task=3, Evaluation protocol=log-probability evaluation2026.06 | 55.3 | 53.8 | 0.032 | |
| SAEBackbone=Gemma-2 9B-it, Epochs per task=3, Evaluation protocol=log-probability evaluation2026.06 | 51 | 49.5 | 0.121 | |
| MASBackbone=Gemma-2 9B-it, Epochs per task=3, Evaluation protocol=log-probability evaluation2026.06 | 45.6 | 44.1 | 0.143 | |
| EWCBackbone=Gemma-2 9B-it, Epochs per task=3, Evaluation protocol=log-probability evaluation2026.06 | 42.3 | 40.9 | 0.165 | |
| UnprotectedBackbone=Gemma-2 9B-it, Epochs per task=3, Evaluation protocol=log-probability evaluation2026.06 | 39 | 37.6 | 0.172 | |
| ReplayReplay buffer size=2%, Backbone=Gemma-2 9B-it, Epochs per task=3, Evaluation protocol=log-probability evaluation2026.06 | 28.7 | 27.4 | 0.056 |