da-discourse-coherence-base

This model is a fine-tuned version of NbAiLab/nb-bert-base on the DDisco dataset. It achieves the following results on the evaluation set:

• Loss: 0.7487
• Accuracy: 0.6915

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 703
• gradient_accumulation_steps: 16
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 6.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
1.3422	0.4	5	1.0166	0.5721
0.9645	0.8	10	0.8966	0.5721
0.9854	1.24	15	0.8499	0.5721
0.8628	1.64	20	0.8379	0.6517
0.9046	2.08	25	0.8228	0.5721
0.8361	2.48	30	0.7980	0.5821
0.8158	2.88	35	0.8095	0.5821
0.8689	3.32	40	0.7989	0.6169
0.8125	3.72	45	0.7730	0.6965
0.843	4.16	50	0.7566	0.6418
0.7421	4.56	55	0.7840	0.6517
0.7949	4.96	60	0.7531	0.6915
0.828	5.4	65	0.7464	0.6816
0.7438	5.8	70	0.7487	0.6915

Framework versions

• Transformers 4.26.0
• Pytorch 1.13.0a0+d0d6b1f
• Datasets 2.9.0
• Tokenizers 0.13.2

Contributor

ajders