Continual Knowledge Learning on LAMA-CKL (Llama2-7B based test)
42.9Top AccuracyTAALM
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| TAALMBackbone=Llama2-7B, PEFT=QLoRA, Epoch=42024.07 | 42.9 | 89.83 | 132.73 | |
| RHO-1Backbone=Llama2-7B, PEFT=QLoRA, Epoch=182024.07 | 14.1 | 82.23 | 96.33 | |
| K-AdapterBackbone=Llama2-7B, PEFT=K-Adapter, Epoch=302024.07 | 12.6 | 99.8 | 112.4 | |
| FinetuneBackbone=Llama2-7B, PEFT=QLoRA, Epoch=162024.07 | 11.5 | 81.74 | 93.24 | |
| RecAdamBackbone=Llama2-7B, PEFT=QLoRA, Epoch=242024.07 | 10 | 79.33 | 89.33 | |
| Mix-reviewBackbone=Llama2-7B, PEFT=QLoRA, Epoch=252024.07 | 8 | 79.88 | 87.88 |