learning rate: 3e-5 | |
training epochs: 5 | |
batch size: 8 | |
seed: 42 | |
model: bert-base-uncased | |
The model is pretrained on MNLI (we use kangnichaluo/mnli-2 directly) and then finetuned on CB which is converted into two-way nli classification (predict entailment or not-entailment class) |