đ Indo-roberta-indonli
Indo-roberta-indonli is a natural language inference classifier based on the Indo-roberta model. It addresses the need for accurate natural language inference in the Indonesian context, leveraging the pre - trained Indo - roberta model and fine - tuning it on relevant datasets.
đ Quick Start
Indo-roberta-indonli is a powerful tool for natural language inference. It uses the Indo-roberta model and is fine - tuned on the IndoNLI dataset. The model was tested on validation, test_layer, and test_expert datasets from the GitHub repository, and the results are presented below.
⨠Features
- Based on Indo - roberta: Utilizes the pre - trained Indo - roberta model for effective transfer learning.
- Trained on IndoNLI: Fine - tuned on the IndoNLI dataset for natural language inference tasks.
đĻ Installation
No specific installation steps are provided in the original document, so this section is skipped.
đģ Usage Examples
Basic Usage
from transformers import pipeline
pretrained_name = "StevenLimcorn/indonesian-roberta-indonli"
nlp = pipeline(
"zero-shot-classification",
model=pretrained_name,
tokenizer=pretrained_name
)
nlp("Amir Sjarifoeddin Harahap lahir di Kota Medan, Sumatera Utara, 27 April 1907. Ia meninggal di Surakarta, Jawa Tengah, pada 19 Desember 1948 dalam usia 41 tahun. </s></s> Amir Sjarifoeddin Harahap masih hidup.")
đ Documentation
Result
Dataset |
Accuracy |
F1 |
Precision |
Recall |
Test Lay |
0.74329 |
0.74075 |
0.74283 |
0.74133 |
Test Expert |
0.6115 |
0.60543 |
0.63924 |
0.61742 |
Model
The model was trained for 5 epochs, with a batch size of 16, a learning rate of 2e - 5, and weight decay of 0.01. The achieved metrics are as follows:
Epoch |
Training Loss |
Validation Loss |
Accuracy |
F1 |
Precision |
Recall |
1 |
0.942500 |
0.658559 |
0.737369 |
0.735552 |
0.735488 |
0.736679 |
2 |
0.649200 |
0.645290 |
0.761493 |
0.759593 |
0.762784 |
0.759642 |
3 |
0.437100 |
0.667163 |
0.766045 |
0.763979 |
0.765740 |
0.763792 |
4 |
0.282000 |
0.786683 |
0.764679 |
0.761802 |
0.762011 |
0.761684 |
5 |
0.193500 |
0.925717 |
0.765134 |
0.763127 |
0.763560 |
0.763489 |
đ§ Technical Details
The model is a transfer - learned natural language inference classifier. It starts with the Indo - roberta model and is fine - tuned on the IndoNLI dataset. The training process involves 5 epochs, a batch size of 16, a learning rate of 2e - 5, and weight decay of 0.01.
đ License
The model is released under the MIT license.
đ Reference
The dataset used is from IndoNLI.
@inproceedings{indonli,
title = "IndoNLI: A Natural Language Inference Dataset for Indonesian",
author = "Mahendra, Rahmad and Aji, Alham Fikri and Louvan, Samuel and Rahman, Fahrurrozi and Vania, Clara",
booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2021",
publisher = "Association for Computational Linguistics",
}
â ī¸ Important Note
Consider the biases that may come from both the pre - trained RoBERTa model and the INDONLI
dataset, as they may affect the model's results.
đ¨âđģ Author
Indonesian RoBERTa Base IndoNLI was trained and evaluated by Steven Limcorn. All computation and development were done on Google Colaboratory using their free GPU access.