w11wo/indonesian-roberta-base-indonli

Indonesian RoBERTa Base IndoNLI

Indonesian RoBERTa Base IndoNLI is a natural language inference (NLI) model based on the RoBERTa model. The model was originally the pre-trained Indonesian RoBERTa Base model, which is then fine-tuned on IndoNLI's dataset consisting of Indonesian Wikipedia, news, and Web articles [1].

After training, the model achieved an evaluation/dev accuracy of 77.06%. On the benchmark test_lay subset, the model achieved an accuracy of 74.24% and on the benchmark test_expert subset, the model achieved an accuracy of 61.66%.

Hugging Face's Trainer class from the Transformers library was used to train the model. PyTorch was used as the backend framework during training, but the model remains compatible with other frameworks nonetheless.

Model

Model	#params	Arch.	Training/Validation data (text)
indonesian-roberta-base-indonli	124M	RoBERTa Base	IndoNLI

Evaluation Results

The model was trained for 5 epochs, with a batch size of 16, a learning rate of 2e-5, a weight decay of 0.1, and a warmup ratio of 0.2, with linear annealing to 0. The best model was loaded at the end.

Epoch	Training Loss	Validation Loss	Accuracy
1	0.989200	0.691663	0.731452
2	0.673000	0.621913	0.766045
3	0.449900	0.662543	0.770596
4	0.293600	0.777059	0.768320
5	0.194200	0.948068	0.764224

How to Use

As NLI Classifier

from transformers import pipeline

pretrained_name = "w11wo/indonesian-roberta-base-indonli"

nlp = pipeline(
    "sentiment-analysis",
    model=pretrained_name,
    tokenizer=pretrained_name
)

nlp("Andi tersenyum karena mendapat hasil baik. </s></s> Andi sedih.")

Disclaimer

Do consider the biases which come from both the pre-trained RoBERTa model and the IndoNLI dataset that may be carried over into the results of this model.

References

[1] Mahendra, R., Aji, A. F., Louvan, S., Rahman, F., & Vania, C. (2021, November). IndoNLI: A Natural Language Inference Dataset for Indonesian. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.

Author

Indonesian RoBERTa Base IndoNLI was trained and evaluated by Wilson Wongso. All computation and development are done on Google Colaboratory using their free GPU access.