deberta-4-disfluency

This model is a fine-tuned version of microsoft/deberta-v3-base on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0093
• Precision: 0.9149
• Recall: 0.9306
• F1: 0.9226
• Accuracy: 0.9841
• Disfluent Precision: 0.9533
• Disfluent Recall: 0.9667
• Disfluent F1: 0.9599
• False Positive Rate: 0.0115

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: 2e-05
• train_batch_size: 64
• eval_batch_size: 192
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy	Disfluent Precision	Disfluent Recall	Disfluent F1	False Positive Rate
0.0079	0.1821	250	0.0090	0.9100	0.9259	0.9179	0.9833	0.9522	0.9633	0.9577	0.0117
0.0077	0.3642	500	0.0098	0.9118	0.9296	0.9206	0.9840	0.9595	0.9592	0.9594	0.0098
0.0059	0.5462	750	0.0146	0.9168	0.9287	0.9227	0.9836	0.9654	0.9508	0.9581	0.0083
0.0058	0.7283	1000	0.0109	0.9157	0.9242	0.9199	0.9837	0.9549	0.9623	0.9586	0.0110
0.0055	0.9104	1250	0.0116	0.9131	0.9279	0.9205	0.9834	0.9530	0.9632	0.9581	0.0115
0.0072	1.0925	1500	0.0093	0.8936	0.9201	0.9067	0.9815	0.9396	0.9679	0.9535	0.0150
0.0062	1.2746	1750	0.0103	0.9085	0.9226	0.9155	0.9822	0.9455	0.9652	0.9553	0.0134
0.0059	1.4567	2000	0.0088	0.9079	0.9225	0.9151	0.9816	0.9371	0.9715	0.9540	0.0157
0.0062	1.6387	2250	0.0093	0.9149	0.9306	0.9226	0.9841	0.9533	0.9667	0.9599	0.0115

Framework versions

• Transformers 4.52.4
• Pytorch 2.6.0+cu124
• Datasets 2.14.4
• Tokenizers 0.21.1