[dataset] image-text datasets
Collection
8 items • Updated
Datasets:
Dataset Overview
This is the filtered Vietnamese subset of the DataComp Large Pool.
The table below shows the processing steps applied to achieve this subset.
| Processing step | # Rows | % |
|---|---|---|
| DataComp Large Pool | 1,280,000,000 | - |
| Filtered for Vietnamese text using the fasttext-language-identification model (score threshold > 0.7) | 26,537,765 | 100% |
| Removed samples with captions <= 10 characters or <= 3 words | 9,451,518 | 35.6% |
| Removed images with a smaller dimension below 200 pixels | 6,817,062 | 25.7% |
| Removed images with an aspect ratio >= 3 | 6,793,921 | 25.6% |
| Number of samples with an mclip_score | 3,950,377 | 14.9% |
The following table shows the percentiles based on the mclip_score computed using clip-ViT-B-32-multilingual-v1.
| Percentile | mclip_score |
|---|---|
| 5th | 0.20242923 |
| 10th | 0.21163453 |
| 15th | 0.21813008 |
| 20th | 0.22349254 |
| 25th | 0.22828135 |
| 30th | 0.23274314 |
| 40th | 0.24120911 |
| 50th | 0.24967791 |
| 60th | 0.25872927 |
| 75th | 0.27487311 |
| 85th | 0.28946092 |
| 90th | 0.29953519 |
| 95th | 0.31465694 |
- Downloads last month
- 21