Table 4. Results of training and validation for several BERT classifiers (damage vs non-damage).

Sequence length	Batch size	Epochs	Learning rate	Epsilon	Training loss	Validation loss	Validity accuracy	Validating F1-score	Training time	Validation time
BERT-64	32	1	2e−5	1e−8	0.07	0.12	0.9094	0.8244	0:02:11	0:00:05
		2	2e−5	1e−8	0.09	0.24	0.9120	0.8928	0:01:58	0:00:04
		3	2e−5	1e−8	0.07	0.37	0.9120	0.8928	0:01:58	0:00:04
		4	2e−5	1e−8	0.03	0.39	0.9055	0.8262	0:01:58	0:00:04
BERT-128	32	1	2e−5	1e−8	0.29	0.23	0.9245	0.8479	0:03:57	0:00:09
		2	2e−5	1e−8	0.19	0.25	0.9333	0.8633	0:03:56	0:00:09
		3	2e−5	1e−8	0.12	0.29	0.9245	0.8479	0:03:56	0:00:09
		4	2e−5	1e−8	0.08	0.33	0.9245	0.8479	0:03:55	0:00:09