AIAT/Pangpuriye-public_ThaiSum40k

🤖 Super AI Engineer Development Program Season 4 - Pangpuriye House - ThaiSum40k

Original Dataset

We adopt this from ThaiSum dataset from https://huggingface.co/datasets/thaisum the original repository. We used this dataset during the fine-tuning of Panguriye's LLM. The dataset is available under the Creative Commons Attribution 2.0. The original dataset consists of 380,868 rows of title, body, and summary in Thai. We modified this dataset by subsetting only 40,000 rows and used จงสรุปเรื่องต่อไปนี้ as an instruction. The input we set by the given context, and the output is the summarization version of the given context.

We think that the ThaiSum dataset will aid in fine-tuning our instruction-tuned LLM in shortening the answers and defining logical summarizations. During the fine-tuning phase, we want to include summaries to reduce the output length as much as feasible.

Call Dataset

The following code is an example calling from datasets library.

from datasets import load_dataset

dataset = load_dataset("AIAT/Pangpuriye-public_ThaiSum40k")

Citation Information

We acknowledge the original dataset, and please redirect to the original paper as follow:

Please refer to the original dataset here https://huggingface.co/datasets/thaisum.

@mastersthesis{chumpolsathien_2020,
    title={Using Knowledge Distillation from Keyword Extraction to Improve the Informativeness of Neural Cross-lingual Summarization},
    author={Chumpolsathien, Nakhun},
    year={2020},
    school={Beijing Institute of Technology}