bert_uncased_L-4_H-256_A-4_cola

This model is a fine-tuned version of google/bert_uncased_L-4_H-256_A-4 on the GLUE COLA dataset. It achieves the following results on the evaluation set:

• Loss: 0.5943
• Matthews Correlation: 0.2651
• Accuracy: 0.7028

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 256
• eval_batch_size: 256
• seed: 10
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 50

Training results

Training Loss	Epoch	Step	Validation Loss	Matthews Correlation	Accuracy
0.6358	1.0	34	0.6182	0.0	0.6913
0.6077	2.0	68	0.6184	0.0	0.6913
0.5982	3.0	102	0.6035	0.0	0.6913
0.575	4.0	136	0.5997	0.1458	0.7009
0.5391	5.0	170	0.5992	0.2018	0.7028
0.4999	6.0	204	0.6159	0.2088	0.7085
0.4722	7.0	238	0.5974	0.2782	0.7248
0.4437	8.0	272	0.5943	0.2651	0.7028
0.4204	9.0	306	0.6239	0.2618	0.7210
0.3956	10.0	340	0.6360	0.2655	0.7191
0.3671	11.0	374	0.6876	0.2592	0.7200
0.3546	12.0	408	0.7041	0.2665	0.7239
0.333	13.0	442	0.6849	0.2891	0.7229

Framework versions

• Transformers 4.46.3
• Pytorch 2.2.1+cu118
• Datasets 2.17.0
• Tokenizers 0.20.3