We utilize a combination of datasets that contain samples in Vietnamese language, ensuring a robust and representative text corpus. These datasets include:

The Vietnamese subset of the C4 dataset .
The Vietnamese subset of the OSCAR dataset, version 23.01.
Wikipedia's Vietnamese articles.
Binhvq's Vietnamese news corpus.

Preprocessing

We use NeMo Curator to curate the collected data. The data curation pipeline includes these key steps:

Unicode Reformatting: Texts are standardized into a consistent Unicode format to avoid encoding issues.
Exact Deduplication: Removes exact duplicates to reduce redundancy.
Quality Filtering:
Heuristic Filtering: Applies rules-based filters to remove low-quality content.
Classifier-Based Filtering: Uses machine learning to classify and filter documents based on quality.

Notebook

Dataset Statistics

Content diversity Domain proportion in curated dataset

Character based metrics Box plots of percentage of symbols, numbers, and whitespace characters compared to the total characters, word counts and average word lengths

Token count distribution Distribution of document sizes (in terms of token count)

Embedding visualization UMAP visualization of 5% of the dataset

Downloads last month: 732

Total file size:

34.6 GB

Models trained or fine-tuned on VTSNLP/vietnamese_curated_dataset