mcemilg commited on
Commit
fec8f5b
1 Parent(s): 97fc3b6

Update Readme.md

Browse files
Files changed (1) hide show
  1. README.md +44 -2
README.md CHANGED
@@ -23,8 +23,50 @@ task_categories:
23
 
24
  - **Homepage:** [laion-5b](https://laion.ai/blog/laion-5b/)
25
  - **Huggingface:** [laion/laion2B-multi](https://huggingface.co/datasets/laion/laion2B-multi)
26
- - **Point of Contact: [mcg@mcemilg.dev](mailto:mcg@mcemilg.dev)**
27
 
28
  ### Dataset Summary
29
 
30
- [LAION-5B](https://laion.ai/blog/laion-5b/) is a large scale openly accessible image-text dataset contains text from multiple languages. This is a Turkish subset data of [laion/laion2B-multi](https://huggingface.co/datasets/laion/laion2B-multi).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
23
 
24
  - **Homepage:** [laion-5b](https://laion.ai/blog/laion-5b/)
25
  - **Huggingface:** [laion/laion2B-multi](https://huggingface.co/datasets/laion/laion2B-multi)
26
+ - **Point of Contact:** [mcemilg](mailto:mcg@mcemilg.dev)
27
 
28
  ### Dataset Summary
29
 
30
+ [LAION-5B](https://laion.ai/blog/laion-5b/) is a large scale openly accessible image-text dataset contains text from multiple languages. This is a Turkish subset data of [laion/laion2B-multi](https://huggingface.co/datasets/laion/laion2B-multi). It's compatible to be used with [image2dataset](https://github.com/rom1504/img2dataset) to fetch the images at scale.
31
+
32
+
33
+ ### Data Structure
34
+
35
+ ```python
36
+ DatasetDict({
37
+ train: Dataset({
38
+ features: ['SAMPLE_ID', 'URL', 'TEXT', 'HEIGHT', 'WIDTH', 'LICENSE', 'LANGUAGE', 'NSFW', 'similarity'],
39
+ num_rows: 34638627
40
+ })
41
+ })
42
+ ```
43
+
44
+ ```python
45
+ {
46
+ 'SAMPLE_ID': Value(dtype='int64', id=None),
47
+ 'URL': Value(dtype='string', id=None),
48
+ 'TEXT': Value(dtype='string', id=None),
49
+ 'HEIGHT': Value(dtype='int64', id=None),
50
+ 'WIDTH': Value(dtype='int64', id=None),
51
+ 'LICENSE': Value(dtype='string', id=None),
52
+ 'LANGUAGE': Value(dtype='string', id=None),
53
+ 'NSFW': Value(dtype='string', id=None),
54
+ 'similarity': Value(dtype='float64', id=None)
55
+ }
56
+ ```
57
+
58
+
59
+ ### Notes
60
+
61
+ The data was basically processed to drop non-Turkish and irrelevant texts before published. Both [FastText](https://fasttext.cc/docs/en/language-identification.html) and [langdetect](https://pypi.org/project/langdetect/) libraries were used to identify if the text is Turkish or not. The cleaning process can be summarized as follows:
62
+
63
+ - replace \"\"\" with empty str
64
+ - remove URLs in texts
65
+ - Drop if both FastText and LangDetect are highly confident with there is no Turkish in text.
66
+ - Drop empty text fields.
67
+
68
+ ### License
69
+ CC-BY-4.0
70
+
71
+
72
+