The model was originally HuggingFace's pretrained English GPT-2 model and is later fine-tuned on the Indonesian dataset. Many of the techniques used are based on a notebook/blog shared by Pierre Guillou, where Pierre Guillou fine-tuned the English GPT-2 model on a Portuguese dataset.

Frameworks used include HuggingFace's Transformers and fast.ai's Deep Learning library. 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)
`indo-gpt2-small`	124M	GPT-2 Small	Indonesian Wikipedia (3.1 GB of text)

Evaluation Results

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

epoch	train loss	valid loss	perplexity	total time
0	2.981	2.936	18.85	2:45:25

How to Use (PyTorch)

Load Model and Byte-level Tokenizer

python

from transformers import GPT2TokenizerFast, GPT2LMHeadModel
pretrained_name = "w11wo/indo-gpt2-small"
tokenizer = GPT2TokenizerFast.from_pretrained(pretrained_name)
tokenizer.model_max_length = 1024
model = GPT2LMHeadModel.from_pretrained(pretrained_name)

Generate a Sequence

python

# sample prompt
prompt = "Nama saya Budi, dari Indonesia"
input_ids = tokenizer.encode(prompt, return_tensors='pt')
model.eval()

# generate output using top-k sampling
sample_outputs = model.generate(input_ids, 
                                pad_token_id=50256,
                                do_sample=True, 
                                max_length=40, 
                                min_length=40,
                                top_k=40,
                                num_return_sequences=1)

for i, sample_output in enumerate(sample_outputs):
    print(tokenizer.decode(sample_output.tolist()))

Disclaimer

Do remember that although the dataset originated from Wikipedia, the model may not always generate factual texts. Additionally, the biases which came from the Wikipedia articles may be carried over into the results of this model.

Credits

Major thanks to Pierre Guillou for sharing his work, which did not only enable me to realize this project but also taught me tons of new, exciting stuff.

Author

Indo GPT-2 Small was trained and evaluated by Wilson Wongso. All computation and development are done on Google Colaboratory using their free GPU access.