Datasets:

Modalities:
Image
Text
Formats:
parquet
ArXiv:
Libraries:
Datasets
Dask
License:
File size: 2,669 Bytes
52dd183
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4ad3e89
 
 
 
 
 
 
 
 
31f6e93
4ad3e89
 
 
 
 
 
 
 
a2dc84f
 
 
 
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
---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: caption
    sequence: string
  - name: img_url
    dtype: string
  - name: category
    dtype: string
  splits:
  - name: train
    num_bytes: 109118946634.129
    num_examples: 1273229
  download_size: 108981536155
  dataset_size: 109118946634.129
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# SEA-VL: A Multicultural Vision-Language Dataset for Southeast Asia

Paper: [Crowdsource, Crawl, or Generate? Creating SEA-VL, A Multicultural Vision-Language Dataset for Southeast Asia]()

Dataset: [SEA-VL Collection on HuggingFace](https://huggingface.co/collections/SEACrowd/sea-vl-multicultural-vl-dataset-for-southeast-asia-67cf223d0c341d4ba2b236e7)

Code: [SEA-VL Experiemnt](https://github.com/SEACrowd/sea-vl-experiments) | [SEA-VL Image Collection](https://github.com/SEACrowd/sea-vl-image-collection)

## What is SEA-VL?
Following the success of our SEACrowd project, we’re excited to announce [SEA-VL](https://seacrowd.github.io/seavl-launch/), a new open-source initiative to create high-quality vision-language datasets specifically for Southeast Asian (SEA) languages! We’re calling on contributors to help us build a SEA-specific vision-language model.

SEA-VL is a big initiative, so we have decided to split it into two phases. In Phase 1 of SEA-VL, we’re looking for self-taken, culturally-relevant images with descriptions about the shared image. This will be cleaned and compiled into a comprehensive open-access SEA-relevant image dataset. To further extend the size of the culturally-relevant images, we also conduct SEA-relevant image dataset collection by developing a robust pipeline for crawling, filtering, and deduplication. This dataset will serve as the foundation for Phase 2, where we’ll develop instruction-tuning VL datasets and build a SEA-specific vision language model (VLM) using the constructed dataset.

The Phase 1 of SEA-VL has been concluded in early 2025. From this initiative, we release [SEA-VL collection](https://huggingface.co/collections/SEACrowd/sea-vl-multicultural-vl-dataset-for-southeast-asia-67cf223d0c341d4ba2b236e7) which covers [SEA-VL Crowdsourcing Dataset (8k samples)](https://huggingface.co/datasets/SEACrowd/sea-vl_crowdsourced) and [SEA-VL Crawling Dataset (1.2M samples)](https://huggingface.co/datasets/SEACrowd/sea-vl_crawling).

If you are interested in joining our initiative, please consider joining our [Discord Server](https://discord.gg/QV2rEyqwQk) and join our regular town hall meeting (More info in Discord)!

## Citing SEA-VL
```
TODO
```