File size: 905 Bytes
c871a8b
8416f3d
c871a8b
8416f3d
c871a8b
8416f3d
c871a8b
8416f3d
c871a8b
 
cb4c129
c871a8b
fb2083d
38e8edd
cb4c129
c871a8b
 
 
 
 
 
38e8edd
c871a8b
 
 
 
38e8edd
c871a8b
 
38e8edd
 
 
 
c871a8b
 
38e8edd
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
---
pretty_name: MMCS-773K
task_categories:
- image-text-to-text
language:
- en
size_categories:
- 100K<n<1M
---

# MMCS Pretraining Data

This repository contains the generated annotations used by [Multi-Modal Code-Switching](https://arxiv.org/abs/2608.11167). Source images are not redistributed here.
The dataset contains **773,779** samples.
Six conversations format JSON files are provided in `annotations/`

| Source | Samples |
| --- | ---: |
| COCO 2014 | 75,033 |
| Flickr30K | 29,140 |
| GQA | 108,516 |
| Objects365 | 342,746 |
| Open Images | 127,237 |
| SA-1B | 91,107 |
| **Total** | **773,779** |

## Download

```bash
hf download LockOnN/MMCS-Data \
    --repo-type dataset \
    --include "annotations/*" \
    --local-dir MMCS-Data
```

See the [MM-CodeSwitch README](https://github.com/Changhao-Xiang/MM-CodeSwitch#data) for the complete data preparation and reproduction instructions.