This repository features a fine-tuned Pegasus X model designed for summarizing Thai text. The architecture of the model is based on the Pegasus X model.

Library

pip install transformers

Example

from transformers import PegasusXForConditionalGeneration, AutoTokenizer

model = PegasusXForConditionalGeneration.from_pretrained("satjawat/pegasus-x-thai-sum")
tokenizer = AutoTokenizer.from_pretrained("satjawat/pegasus-x-thai-sum")

new_input_string = "ข้อความ"
new_input_ids = tokenizer(new_input_string.lower(), return_tensors="pt").input_ids
summary_ids = model.generate(new_input_ids, max_length=50, num_beams=6, length_penalty=2.0, early_stopping=True)
summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)

print("Input:", new_input_string)
print("Generated Summary:", summary)

Training hyperparameters

The following hyperparameters were used during training:

accumulation_steps: 2
num_epochs: 20
num_beams: 6
learning_rate: lr=5e-5
optimizer: AdamW with betas=(0.9,0.999) and epsilon=1e-08
activation_function: gelu
add_bias_logits: True
normalize_embedding: True
add_final_layer_norm: False
normalize_before: False

Score

Evaluate the model with the test dataset of ThaiSum, consisting of a total of 11,000 articles, with the following scores:

Rouge1: 0.490279
Rouge2: 0.289839
Rougel: 0.489334

Resource Funding

NSTDA Supercomputer center (ThaiSC) and the National e-Science Infrastructure Consortium for their support of computer facilities.

Citation

If you use "satjawat/pegasus-x-thai-sum" in your project or publication, please cite the model as follows:

ปรีชานนท์ ชาติไทย และ สัจจวัจน์ ส่งเสริม. (2567),
การสรุปข้อความข่าวภาษาไทยด้วยโครงข่ายประสาทเทียม (Thai News Text Summarization Using Neural Network),
วิทยาศาสตรบัณฑิต (วทบ.):ขอนแก่น, มหาวิทยาลัยขอนแก่น