mdeberta-v3-base-azsci-topics

This model is a fine-tuned version of microsoft/mdeberta-v3-base on hajili/azsci_topics dataset. It achieves the following results on the evaluation set:

• Loss: 0.5213
• Precision: 0.8633
• Recall: 0.8759
• F1: 0.8685
• Accuracy: 0.8759

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
No log	1.0	288	0.9919	0.6713	0.7465	0.6959	0.7465
1.4813	2.0	576	0.7035	0.7994	0.8238	0.8022	0.8238
1.4813	3.0	864	0.5605	0.8540	0.8568	0.8462	0.8568
0.5512	4.0	1152	0.5296	0.8615	0.8689	0.8623	0.8689
0.5512	5.0	1440	0.5213	0.8633	0.8759	0.8685	0.8759

Evaluation results

Topic	Precision	Recall	F1	Support
Aqrar elmlər	0.678571	0.703704	0.690909	27
Astronomiya	0	0	0	2
Biologiya elmləri	0.877358	0.885714	0.881517	105
Coğrafiya	0.833333	0.882353	0.857143	17
Filologiya elmləri	0.932203	0.942857	0.9375	175
Fizika	0.763158	0.852941	0.805556	34
Fəlsəfə	0.294118	0.357143	0.322581	14
Hüquq elmləri	0.965517	0.965517	0.965517	29
Kimya	0.828571	0.95082	0.885496	61
Memarlıq	0	0	0	5
Mexanika	0	0	0	4
Pedaqogika	0.882353	0.957447	0.918367	47
Psixologiya	1	0.722222	0.83871	18
Riyaziyyat	0.871795	0.871795	0.871795	39
Siyasi elmlər	0.807692	0.84	0.823529	25
Sosiologiya	0	0	0	4
Sənətşünaslıq	0.82	0.87234	0.845361	47
Tarix	0.846154	0.846154	0.846154	78
Texnika elmləri	0.822917	0.759615	0.79	104
Tibb elmləri	0.953947	0.97973	0.966667	148
Yer elmləri	0.7	0.538462	0.608696	13
İqtisad elmləri	0.948052	0.960526	0.954248	152
Əczaçılıq elmləri	0	0	0	4
macro avg	0.644597	0.647363	0.643902	1152
weighted avg	0.863306	0.875868	0.868519	1152

Framework versions

• Transformers 4.38.2
• Pytorch 2.1.0+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2