I am unable to reproduce the same results over different model runs.
When using bert-base-uncased , the generated results (and loss values) will be exactly the same across runs. However, when using allenai/longformer-base-4096 results (and loss values) will vary across runs.
Please help me. This is urgent.