Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,36 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Altogether-FT
|
2 |
+
|
3 |
+
(EMNLP 2024) Altogether-FT is an annotated fine-tuning dataset that re-aligns alt-texts into dense captions. It powers altogether captioner to transform Internet-scale quality alt-texts into dense captions, instead of captioning from scratch as naive captions (e.g, "a dog is walking in the park.").
|
4 |
+
It contains 15448 examples for training and 500 examples for evaluation from WIT and DataComp.
|
5 |
+
|
6 |
+
![Altogether](altogether.png)
|
7 |
+
|
8 |
+
|
9 |
+
```bibtex
|
10 |
+
@inproceedings{xu2024altogether,
|
11 |
+
title={Altogether: Image Captioning via Re-aligning Alt-text},
|
12 |
+
author={Hu Xu, Po-Yao Huang, Xiaoqing Ellen Tan, Ching-Feng Yeh, Jacob Kahn, Christine Jou, Gargi Ghosh, Omer Levy, Luke Zettlemoyer, Wen-tau Yih, Shang-Wen Li, Saining Xie and Christoph Feichtenhofer},
|
13 |
+
journal={arXiv preprint arXiv:xxxx.xxxxx},
|
14 |
+
year={2024}
|
15 |
+
}
|
16 |
+
```
|
17 |
+
|
18 |
+
## Altogether-FT
|
19 |
+
|
20 |
+
```python
|
21 |
+
from datasets import load_dataset
|
22 |
+
|
23 |
+
train_dataset = load_dataset("json", data_files="activebus/Altogether-FT/altogether_ft_train.json", field="data")
|
24 |
+
|
25 |
+
eval_dataset = load_dataset("json", data_files="activebus/Altogether-FT/altogether_ft_eval.json", field="data")
|
26 |
+
```
|
27 |
+
|
28 |
+
|
29 |
+
## License
|
30 |
+
|
31 |
+
The majority of Altogether-FT is licensed under CC-BY-NC, portions of the project are available under separate license terms: CLIPCap is licensed MIT and open_clip is licensed under the https://github.com/mlfoundations/open_clip license.
|
32 |
+
|
33 |
+
|
34 |
+
---
|
35 |
+
license: cc-by-nc-4.0
|
36 |
+
---
|