search

sail-sg / CLoT

CVPR'24, Official Codebase of our Paper: "Let's Think Outside the Box: Exploring Leap-of-Thought in Large Language Models with Creative Humor Generation".
https://zhongshsh.github.io/CLoT
296 stars 12 forks source link

In jp.jsonl, all the question is "null", is there any mistake? #5

Open Labmem009 opened 6 months ago

Labmem009 commented 6 months ago

Some lines are "T2T" type, but still have no question, is there any mistake?

zhongshsh commented 6 months ago

Not a mistake. Some T2T type questions have placed text on images, for example,

Thank you for the reminder, we've added clarification to the dataset card.

Labmem009 commented 6 months ago

Not a mistake. Some T2T type questions have placed text on images, for example, image

But all the "T2T" lines have "null" question, I checked that there is no line has question text

zhongshsh commented 6 months ago

This is because Japanese data mainly comes from Bokete, where T2T is displayed in image form.

This type of data can also be labeled as T2T because, as shown in the following image (chat with Qwen-VL), well-formatted text within images can be easily recognized and understood by multimodal large language models.

You also can find data where the question is not 'None' in the Chinese and English data.