Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
@@ -1,117 +1,117 @@
|
|
1 |
---
|
|
|
2 |
language:
|
3 |
- en
|
4 |
-
license: mit
|
5 |
-
task_categories:
|
6 |
-
- image-text-to-text
|
7 |
-
size_categories:
|
8 |
-
- 1M<n<10M
|
9 |
-
pretty_name: mmE5 labeled data
|
10 |
tags:
|
11 |
- embedding
|
12 |
- multimodal
|
|
|
|
|
|
|
13 |
configs:
|
14 |
- config_name: TAT-DQA
|
15 |
data_files:
|
16 |
-
|
17 |
-
|
18 |
- config_name: ArxivQA
|
19 |
data_files:
|
20 |
-
|
21 |
-
|
22 |
- config_name: InfoSeek_it2t
|
23 |
data_files:
|
24 |
-
|
25 |
-
|
26 |
- config_name: InfoSeek_it2it
|
27 |
data_files:
|
28 |
-
|
29 |
-
|
30 |
- config_name: ImageNet_1K
|
31 |
data_files:
|
32 |
-
|
33 |
-
|
34 |
- config_name: N24News
|
35 |
data_files:
|
36 |
-
|
37 |
-
|
38 |
- config_name: HatefulMemes
|
39 |
data_files:
|
40 |
-
|
41 |
-
|
42 |
- config_name: SUN397
|
43 |
data_files:
|
44 |
-
|
45 |
-
|
46 |
- config_name: VOC2007
|
47 |
data_files:
|
48 |
-
|
49 |
-
|
50 |
- config_name: InfographicsVQA
|
51 |
data_files:
|
52 |
-
|
53 |
-
|
54 |
- config_name: ChartQA
|
55 |
data_files:
|
56 |
-
|
57 |
-
|
58 |
- config_name: A-OKVQA
|
59 |
data_files:
|
60 |
-
|
61 |
-
|
62 |
- config_name: DocVQA
|
63 |
data_files:
|
64 |
-
|
65 |
-
|
66 |
- config_name: OK-VQA
|
67 |
data_files:
|
68 |
-
|
69 |
-
|
70 |
- config_name: Visual7W
|
71 |
data_files:
|
72 |
-
|
73 |
-
|
74 |
- config_name: VisDial
|
75 |
data_files:
|
76 |
-
|
77 |
-
|
78 |
- config_name: CIRR
|
79 |
data_files:
|
80 |
-
|
81 |
-
|
82 |
- config_name: NIGHTS
|
83 |
data_files:
|
84 |
-
|
85 |
-
|
86 |
- config_name: WebQA
|
87 |
data_files:
|
88 |
-
|
89 |
-
|
90 |
- config_name: VisualNews_i2t
|
91 |
data_files:
|
92 |
-
|
93 |
-
|
94 |
- config_name: VisualNews_t2i
|
95 |
data_files:
|
96 |
-
|
97 |
-
|
98 |
- config_name: MSCOCO_i2t
|
99 |
data_files:
|
100 |
-
|
101 |
-
|
102 |
- config_name: MSCOCO_t2i
|
103 |
data_files:
|
104 |
-
|
105 |
-
|
106 |
- config_name: MSCOCO
|
107 |
data_files:
|
108 |
-
|
109 |
-
|
110 |
---
|
111 |
-
|
112 |
# mmE5 Labeled Data
|
113 |
|
114 |
-
|
|
|
|
|
|
|
115 |
- **MMEB** (with hard negative)
|
116 |
- **InfoSeek** (from M-BEIR)
|
117 |
- **TAT-DQA**
|
@@ -125,27 +125,13 @@ First, you should prepare the images used for training:
|
|
125 |
|
126 |
### Image Downloads
|
127 |
|
128 |
-
- **Download
|
129 |
-
- [**MMEB**](https://huggingface.co/datasets/TIGER-Lab/MMEB-train)
|
130 |
-
- [**InfoSeek**](https://huggingface.co/datasets/TIGER-Lab/M-BEIR)
|
131 |
-
- [**ArxivQA**](https://huggingface.co/datasets/MMInstruction/ArxivQA)
|
132 |
-
- [**TAT-DQA**](https://huggingface.co/datasets/vidore/tatdqa_train)
|
133 |
-
|
134 |
-
For TAT-DQA, you need to first save images into the overall image folder to align usage:
|
135 |
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
image_out_dir = "images/TAT-DQA"
|
142 |
-
os.makedirs(image_out_dir, exist_ok=True)
|
143 |
-
for i, sample in enumerate(dataset):
|
144 |
-
save_path = os.path.join(image_out_dir, f"tatdqa_{i}.png")
|
145 |
-
if os.path.exists(save_path):
|
146 |
-
continue
|
147 |
-
image = sample["image"]
|
148 |
-
image.save(save_path, format="PNG")
|
149 |
```
|
150 |
|
151 |
### Image Organization
|
|
|
1 |
---
|
2 |
+
license: mit
|
3 |
language:
|
4 |
- en
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
tags:
|
6 |
- embedding
|
7 |
- multimodal
|
8 |
+
pretty_name: mmE5 labeled data
|
9 |
+
size_categories:
|
10 |
+
- 1M<n<10M
|
11 |
configs:
|
12 |
- config_name: TAT-DQA
|
13 |
data_files:
|
14 |
+
- split: train
|
15 |
+
path: "TAT-DQA/TAT-DQA.parquet"
|
16 |
- config_name: ArxivQA
|
17 |
data_files:
|
18 |
+
- split: train
|
19 |
+
path: "ArxivQA/ArxivQA.parquet"
|
20 |
- config_name: InfoSeek_it2t
|
21 |
data_files:
|
22 |
+
- split: train
|
23 |
+
path: "InfoSeek_it2t/InfoSeek_it2t.parquet"
|
24 |
- config_name: InfoSeek_it2it
|
25 |
data_files:
|
26 |
+
- split: train
|
27 |
+
path: "InfoSeek_it2it/InfoSeek_it2it.parquet"
|
28 |
- config_name: ImageNet_1K
|
29 |
data_files:
|
30 |
+
- split: train
|
31 |
+
path: "ImageNet_1K/ImageNet_1K.parquet"
|
32 |
- config_name: N24News
|
33 |
data_files:
|
34 |
+
- split: train
|
35 |
+
path: "N24News/N24News.parquet"
|
36 |
- config_name: HatefulMemes
|
37 |
data_files:
|
38 |
+
- split: train
|
39 |
+
path: "HatefulMemes/HatefulMemes.parquet"
|
40 |
- config_name: SUN397
|
41 |
data_files:
|
42 |
+
- split: train
|
43 |
+
path: "SUN397/SUN397.parquet"
|
44 |
- config_name: VOC2007
|
45 |
data_files:
|
46 |
+
- split: train
|
47 |
+
path: "VOC2007/VOC2007.parquet"
|
48 |
- config_name: InfographicsVQA
|
49 |
data_files:
|
50 |
+
- split: train
|
51 |
+
path: "InfographicsVQA/InfographicsVQA.parquet"
|
52 |
- config_name: ChartQA
|
53 |
data_files:
|
54 |
+
- split: train
|
55 |
+
path: "ChartQA/ChartQA.parquet"
|
56 |
- config_name: A-OKVQA
|
57 |
data_files:
|
58 |
+
- split: train
|
59 |
+
path: "A-OKVQA/A-OKVQA.parquet"
|
60 |
- config_name: DocVQA
|
61 |
data_files:
|
62 |
+
- split: train
|
63 |
+
path: "DocVQA/DocVQA.parquet"
|
64 |
- config_name: OK-VQA
|
65 |
data_files:
|
66 |
+
- split: train
|
67 |
+
path: "OK-VQA/OK-VQA.parquet"
|
68 |
- config_name: Visual7W
|
69 |
data_files:
|
70 |
+
- split: train
|
71 |
+
path: "Visual7W/Visual7W.parquet"
|
72 |
- config_name: VisDial
|
73 |
data_files:
|
74 |
+
- split: train
|
75 |
+
path: "VisDial/VisDial.parquet"
|
76 |
- config_name: CIRR
|
77 |
data_files:
|
78 |
+
- split: train
|
79 |
+
path: "CIRR/CIRR.parquet"
|
80 |
- config_name: NIGHTS
|
81 |
data_files:
|
82 |
+
- split: train
|
83 |
+
path: "NIGHTS/NIGHTS.parquet"
|
84 |
- config_name: WebQA
|
85 |
data_files:
|
86 |
+
- split: train
|
87 |
+
path: "WebQA/WebQA.parquet"
|
88 |
- config_name: VisualNews_i2t
|
89 |
data_files:
|
90 |
+
- split: train
|
91 |
+
path: "VisualNews_i2t/VisualNews_i2t.parquet"
|
92 |
- config_name: VisualNews_t2i
|
93 |
data_files:
|
94 |
+
- split: train
|
95 |
+
path: "VisualNews_t2i/VisualNews_t2i.parquet"
|
96 |
- config_name: MSCOCO_i2t
|
97 |
data_files:
|
98 |
+
- split: train
|
99 |
+
path: "MSCOCO_i2t/MSCOCO_i2t.parquet"
|
100 |
- config_name: MSCOCO_t2i
|
101 |
data_files:
|
102 |
+
- split: train
|
103 |
+
path: "MSCOCO_t2i/MSCOCO_t2i.parquet"
|
104 |
- config_name: MSCOCO
|
105 |
data_files:
|
106 |
+
- split: train
|
107 |
+
path: "MSCOCO/MSCOCO.parquet"
|
108 |
---
|
|
|
109 |
# mmE5 Labeled Data
|
110 |
|
111 |
+
|
112 |
+
|
113 |
+
|
114 |
+
This dataset contains datasets used for the supervised finetuning of mmE5 ([mmE5: Improving Multimodal Multilingual Embeddings via High-quality Synthetic Data](https://arxiv.org/abs/2502.08468)):
|
115 |
- **MMEB** (with hard negative)
|
116 |
- **InfoSeek** (from M-BEIR)
|
117 |
- **TAT-DQA**
|
|
|
125 |
|
126 |
### Image Downloads
|
127 |
|
128 |
+
- **Download All Images Used in mmE5**:
|
|
|
|
|
|
|
|
|
|
|
|
|
129 |
|
130 |
+
You can use the script provided in our [source code](https://github.com/haon-chen/mmE5) to download all images used in mmE5.
|
131 |
+
```bash
|
132 |
+
git clone https://github.com/haon-chen/mmE5
|
133 |
+
cd mmE5
|
134 |
+
bash scripts/prepare_images.sh
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
135 |
```
|
136 |
|
137 |
### Image Organization
|