File size: 1,997 Bytes
458840f
 
 
 
 
 
 
 
 
 
 
 
d1a88bb
458840f
 
 
 
 
 
 
 
 
 
 
d1a88bb
 
 
 
 
 
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
---
task_categories:
- image-to-text
- text2text-generation
language:
- ja
---

## 読み込み方
```python
from datasets import load_dataset

dataset = load_dataset("YANS-official/ogiri-debug", split="test")
```

## 概要
大喜利生成の動作確認用データセットです。  
以下の3タスクが含まれます。  

* text_to_text: テキストでお題が渡され、それに対する回答を返します。
* image_to_text: いわゆる「画像で一言」です。画像のみが渡されて、テキストによる回答を返します。
* text_image_to_text: 画像中にテキストが書かれています。テキストの一部が空欄になっているので、そこに穴埋めする形で回答を返します。

## データセットの各カラム説明
| カラム名                  | 型    | 例                      | 概要                                                                                                          |
|-----------------------|------|------------------------|-------------------------------------------------------------------------------------------------------------|
| odai_id               | str  | "origi-dummy-1"        | お題のID                                                                                                       |
| file_path             | str  | "dummy.png"            | 画像のファイル名                                                                                                    |
| type                  | str  | "text_to_text"         | "text_to_text", "image_to_text", "image_text_to_text"のどれかが入っている。                                            |
| odai                  | str  | "この学会、なんだか疲れるなぁ…どんなの?" | "text_to_text"の場合はお題。"image_to_text"の場合は"画像で一言"という文字列が格納されている。"image_text_to_text"は画像をOCRした結果に"[空欄]"が入ったもの。 |