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)