polibert-malaysia-ver3

This model is new version of YagiASAFAS/polibert-malaysia-ver2. What is new is that this model used a new dataset which not only used tnwei/ms-newspapers dataset but also almost 10k of instagram posts regarding several topics about Malaysia. By doing so, this model captures not only formal sentences such as News, but also captures informal sentences such as personal posts. As a tradeoff, the accuracy was quite lower compared to the previous one

• Loss: 2.2001
• Accuracy: 0.6961

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: 3e-05
• train_batch_size: 8
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 16
• mixed_precision_training: Native AMP

Label Mappings

• 0: Economic Concerns
• 1: Racial discrimination or polarization
• 2: Leadership weaknesses
• 3: Development and infrastructure gaps
• 4: Corruption
• 5: Political instablility
• 6: Socials and Public safety
• 7: Administration
• 8: Education
• 9: Religion issues
• 10: Environmental

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
0.9899	1.0	2083	0.9351	0.7007
0.8717	2.0	4166	0.9117	0.7144
0.7277	3.0	6249	1.0107	0.7079
0.5764	4.0	8332	1.0944	0.7069
0.4703	5.0	10415	1.3644	0.7067
0.3377	6.0	12498	1.5634	0.7012
0.3188	7.0	14581	1.7475	0.6925
0.26	8.0	16664	1.7831	0.6944
0.206	9.0	18747	1.8693	0.6932
0.1994	10.0	20830	2.0837	0.6894
0.1718	11.0	22913	2.0601	0.6954
0.1635	12.0	24996	2.0624	0.6901
0.153	13.0	27079	2.0744	0.6863
0.1595	14.0	29162	2.1222	0.6925
0.153	15.0	31245	2.1712	0.6988
0.1188	16.0	33328	2.2001	0.6961

Framework versions

• Transformers 4.18.0
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.12.1