w11wo/malaysian-distilbert-small

Malaysian DistilBERT Small

Malaysian DistilBERT Small is a masked language model based on the DistilBERT model. It was trained on the OSCAR dataset, specifically the unshuffled_original_ms subset.

The model was originally HuggingFace's pretrained English DistilBERT model and is later fine-tuned on the Malaysian dataset. It achieved a perplexity of 10.33 on the validation dataset (20% of the dataset). Many of the techniques used are based on a Hugging Face tutorial notebook written by Sylvain Gugger, and fine-tuning tutorial notebook written by Pierre Guillou.

Hugging Face's Transformers library was used to train the model -- utilizing the base DistilBERT model and their Trainer class. PyTorch was used as the backend framework during training, but the model remains compatible with TensorFlow nonetheless.

Model

Model	#params	Arch.	Training/Validation data (text)
malaysian-distilbert-small	66M	DistilBERT Small	OSCAR unshuffled_original_ms Dataset

Evaluation Results

The model was trained for 1 epoch and the following is the final result once the training ended.

train loss	valid loss	perplexity	total time
2.476	2.336	10.33	0:40:05

How to Use

As Masked Language Model

from transformers import pipeline

pretrained_name = "w11wo/malaysian-distilbert-small"

fill_mask = pipeline(
    "fill-mask",
    model=pretrained_name,
    tokenizer=pretrained_name
)

fill_mask("Henry adalah seorang lelaki yang tinggal di [MASK].")

Feature Extraction in PyTorch

from transformers import DistilBertModel, DistilBertTokenizerFast

pretrained_name = "w11wo/malaysian-distilbert-small"
model = DistilBertModel.from_pretrained(pretrained_name)
tokenizer = DistilBertTokenizerFast.from_pretrained(pretrained_name)

prompt = "Bolehkah anda [MASK] Bahasa Melayu?"
encoded_input = tokenizer(prompt, return_tensors='pt')
output = model(**encoded_input)

Disclaimer

Do consider the biases which came from the OSCAR dataset that may be carried over into the results of this model.

Author

Malaysian DistilBERT Small was trained and evaluated by Wilson Wongso. All computation and development are done on Google Colaboratory using their free GPU access.