VTSNLP commited on
Commit
dccb75b
·
verified ·
1 Parent(s): f876788

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +22 -3
README.md CHANGED
@@ -19,10 +19,29 @@ configs:
19
  - split: train
20
  path: data/train-*
21
  ---
 
22
 
23
- This is a Vietnamese Curated Text Dataset.
24
 
25
- This dataset is collected from multiple open Vietnamese datasets, and curated with [NeMo Curator](https://github.com/NVIDIA/NeMo-Curator)
 
26
 
27
 
28
- ![Proportion of domains in the dataset](https://github.com/HoangNV2001/image-bucket/blob/main/datadist_pie.png)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
19
  - split: train
20
  path: data/train-*
21
  ---
22
+ ### Dataset Description
23
 
24
+ Vietnamese Curated Text Dataset. This dataset is collected from multiple open Vietnamese datasets, and curated with [NeMo Curator](https://github.com/NVIDIA/NeMo-Curator)
25
 
26
+ - **Developed by:** Viettel Solution
27
+ - **Language:** Vietnamese
28
 
29
 
30
+
31
+ ### Details
32
+
33
+ #### Data Collection
34
+ We utilize a combination of datasets that contain samples in Vietnamese language, ensuring a robust and representative text corpus. These datasets include:
35
+ - The Vietnamese subset of the [C4 dataset](https://huggingface.co/datasets/allenai/c4/viewer/vi) .
36
+ - The Vietnamese subset of the [OSCAR dataset, version 23.01](https://huggingface.co/datasets/oscar-corpus/OSCAR-2301/tree/main/vi_meta).
37
+ - [Wikipedia's Vietnamese articles](https://huggingface.co/datasets/wikimedia/wikipedia/viewer/20231101.vi).
38
+ - [Binhvq's Vietnamese news corpus](https://huggingface.co/datasets/jetaudio/binhvq_news).
39
+
40
+
41
+ #### Preprocessing
42
+ We use [NeMo Curator](https://github.com/NVIDIA/NeMo-Curator) to curate the collected data. The data curation pipeline includes these key steps:
43
+ 1. Unicode Reformatting: Texts are standardized into a consistent Unicode format to avoid encoding issues.
44
+ 2. Exact Deduplication: Removes exact duplicates to reduce redundancy.
45
+ 3. Quality Filtering:
46
+ 4. Heuristic Filtering: Applies rules-based filters to remove low-quality content.
47
+ 5. Classifier-Based Filtering: Uses machine learning to classify and filter documents based on quality.