linyq commited on
Commit
a9ac106
1 Parent(s): aa8c91b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +37 -1
README.md CHANGED
@@ -1,4 +1,5 @@
1
  ---
 
2
  dataset_info:
3
  features:
4
  - name: SAMPLE_ID
@@ -27,7 +28,7 @@ dataset_info:
27
  - name: train
28
  num_bytes: 28506248899
29
  num_examples: 107166507
30
- download_size: 16353146269
31
  dataset_size: 28506248899
32
  configs:
33
  - config_name: default
@@ -35,3 +36,38 @@ configs:
35
  - split: train
36
  path: data/train-*
37
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: cc-by-4.0
3
  dataset_info:
4
  features:
5
  - name: SAMPLE_ID
 
28
  - name: train
29
  num_bytes: 28506248899
30
  num_examples: 107166507
31
+ download_size: 16353125308
32
  dataset_size: 28506248899
33
  configs:
34
  - config_name: default
 
36
  - split: train
37
  path: data/train-*
38
  ---
39
+ # 100M Text Debiased Subset from LAION 2B
40
+
41
+ - Captions in LAION-2B have a significant bias towards describing visual text content embedded in the images.
42
+ - Released CLIP models have strong text spotting bias in almost every style of web images, resulting in the CLIP-filtering datasets inherently biased towards visual text dominant data.
43
+ - CLIP models easily learn text spotting capacity from parrot captions while failing to connect the vision-language semantics, just like a text spotting parrot.
44
+
45
+ For more details, please see our [paper](https://arxiv.org/abs/2312.14232).
46
+
47
+ ## Filtering Details
48
+
49
+ We provide an alternative solution by releasing a less biased filtered LAION-2B 100M(107,166,507) subset.
50
+
51
+ We construct a less biased 100M subset from the LAION-2B subset with Empty OCR results, CLIP score > 0.3, and Aesthetics score > 4.5.
52
+
53
+ We add the ase_scores and K-means labels (4000 total) for each image-text pair.
54
+
55
+ *We also released the dataset on [OpenDataLab](https://openxlab.org.cn/datasets/opendatalab-linyiqi/LAION-text-debiased-100M).*
56
+
57
+ The pre-trained CLIP model is released on [github](https://github.com/opendatalab/CLIP-Parrot-Bias).
58
+
59
+ ## Reference
60
+ ```
61
+ @article{lin2023parrot,
62
+ title={Parrot Captions Teach CLIP to Spot Text},
63
+ author={Yiqi Lin and Conghui He and Alex Jinpeng Wang and Bin Wang and Weijia Li and Mike Zheng Shou},
64
+ journal={arXiv preprint arXiv:2312.14232},
65
+ year={2023}
66
+ }
67
+ @misc{conghui2022opendatalab,
68
+ author={He, Conghui and Li, Wei and Jin, Zhenjiang and Wang, Bin and Xu, Chao and Lin, Dahua},
69
+ title={OpenDataLab: Empowering General Artificial Intelligence with Open Datasets},
70
+ howpublished = {\url{https://opendatalab.com}},
71
+ year={2022}
72
+ }
73
+ ```