File size: 3,140 Bytes
76bc1a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dccb75b
f876788
dccb75b
f876788
6ffb4f7
dccb75b
f876788
dccb75b
 
b81fcce
 
dccb75b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dea46ed
c569ad6
180d862
dea46ed
 
8d2e7e0
177563c
dea46ed
8d2e7e0
069c524
dea46ed
8d2e7e0
069c524
dea46ed
8d2e7e0
177563c
180d862
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: id
    dtype: string
  - name: domain
    dtype: string
  splits:
  - name: train
    num_bytes: 65506190827
    num_examples: 12169131
  download_size: 34648619492
  dataset_size: 65506190827
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
### Dataset Description

Vietnamese Curated Text Dataset. This dataset is collected from multiple open Vietnamese datasets, and curated with [NeMo Curator](https://github.com/NVIDIA/NeMo-Curator)

- **Developed by:** Viettel Solutions
- **Language:** Vietnamese

### Details

Please visit our Tech Blog post on NVIDIA's plog page for details. [Link](https://developer.nvidia.com/blog/processing-high-quality-vietnamese-language-data-with-nvidia-nemo-curator/)

#### Data Collection
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](https://huggingface.co/datasets/allenai/c4/viewer/vi) .
- The Vietnamese subset of the [OSCAR dataset, version 23.01](https://huggingface.co/datasets/oscar-corpus/OSCAR-2301/tree/main/vi_meta).
- [Wikipedia's Vietnamese articles](https://huggingface.co/datasets/wikimedia/wikipedia/viewer/20231101.vi).
- [Binhvq's Vietnamese news corpus](https://huggingface.co/datasets/jetaudio/binhvq_news).


#### Preprocessing
We use [NeMo Curator](https://github.com/NVIDIA/NeMo-Curator) to curate the collected data. The data curation pipeline includes these key steps:
1. Unicode Reformatting: Texts are standardized into a consistent Unicode format to avoid encoding issues.
2. Exact Deduplication: Removes exact duplicates to reduce redundancy.
3. Quality Filtering:
4. Heuristic Filtering: Applies rules-based filters to remove low-quality content.
5. Classifier-Based Filtering: Uses machine learning to classify and filter documents based on quality.

**[Notebook](https://github.com/NVIDIA/NeMo-Curator/blob/main/tutorials/pretraining-vietnamese-data-curation/pretraining-vietnamese-data-curation.ipynb)**

#### Dataset Statistics

**Content diversity**
<img src="https://cdn-uploads.huggingface.co/production/uploads/661766c00c68b375f3f0ccc3/mW6Pct3uyP_XDdGmE8EP3.png" alt="Domain proportion in curated dataset" width="500"/>

**Character based metrics**
<img src="https://cdn-uploads.huggingface.co/production/uploads/661766c00c68b375f3f0ccc3/W9TQjM2vcC7uXozyERHSQ.png" alt="Box plots of percentage of symbols, numbers, and whitespace characters compared to the total characters, word counts and average word lengths" width="900"/>

**Token count distribution**
<img src="https://cdn-uploads.huggingface.co/production/uploads/661766c00c68b375f3f0ccc3/PDelYpBI0DefSmQgFONgE.png" alt="Distribution of document sizes (in terms of token count)" width="500"/>

**Embedding visualization**
<img src="https://cdn-uploads.huggingface.co/production/uploads/661766c00c68b375f3f0ccc3/sfeoZWuQ7DcSpbmUOJ12r.png" alt="UMAP visualization of 5% of the dataset" width="650"/>
*UMAP visualization of 5% of the dataset*