IndoRetNet-Liputan6

This model is a Indonesian RetNet model train using the Liputan6 dataset. Using Tokenizer from IndoBERT It achieves the following results on the evaluation set:

• Loss: 3.4936

Model description

Demonstrate training and recurrent inference using a retentive network (https://arxiv.org/pdf/2307.08621.pdf). The code utilizes Sehyun Choi's implementation of retentive network (https://github.com/syncdoth/RetNet).

• License: Apache 2.0.

Intended uses & limitations

Intended to demonstrate training and (recurrent O(1)) inference using a retentive network in Indonesian language.

Training and evaluation data

Using Train and validation set from Liputan6 dataset provided by NusaCrowd.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0006
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 3.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
4.5053	0.17	1000	4.5145
4.1281	0.34	2000	4.1702
3.9452	0.52	3000	4.0094
3.8302	0.69	4000	3.8972
3.6955	0.86	5000	3.8144
3.589	1.03	6000	3.7600
3.5279	1.21	7000	3.7088
3.4598	1.38	8000	3.6670
3.4445	1.55	9000	3.6259
3.4098	1.72	10000	3.5904
3.3455	1.9	11000	3.5610
3.2306	2.07	12000	3.5406
3.261	2.24	13000	3.5216
3.2204	2.41	14000	3.5111
3.2321	2.59	15000	3.5001
3.2514	2.76	16000	3.4941
3.233	2.93	17000	3.4936

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.0+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0