[GH-ISSUE #88] About Image Captioning #75

New issue

Open

opened 2026-10-26 21:11:27 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 21:11:27 +00:00

Owner

Originally created by @adrianJW421 on GitHub (Feb 18, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL2/issues/88

看上去训练集里也有captioning任务, 实测感觉captioning基本遵从不了额外指令, 输出比较直白简单, scene中的很多细节没有描述.

我的prompt没有加任何任务tag, 是这个问题么?

Originally created by @adrianJW421 on GitHub (Feb 18, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL2/issues/88 看上去训练集里也有captioning任务, 实测感觉captioning基本遵从不了额外指令, 输出比较直白简单, scene中的很多细节没有描述. 我的prompt没有加任何任务tag, 是这个问题么?

bsakaguchi commented

2026-10-26 21:11:28 +00:00

Author

Owner

@HubHop commented on GitHub (Feb 19, 2025):

Hi @adrianJW421, can you share your prompts in your cases?

@HubHop commented on GitHub (Feb 19, 2025): Hi @adrianJW421, can you share your prompts in your cases?

bsakaguchi commented

2026-10-26 21:11:28 +00:00

Author

Owner

@adrianJW421 commented on GitHub (Feb 19, 2025):

Hi @adrianJW421, can you share your prompts in your cases?

Thx for relying. Currently in my experiment, I used following prompts:

Can you describe the scene in the images, highlighting the locations of obejcts in the scene? For example, X is on Y, X is under Y, X is near Y.

This some times leads to immediate 'EOS'

I also try to improve with <[grounding]>, it did looks better(I mean, improved a bit).

<[grounding]>Can you describe the scene in the images, highlighting the locations of obejcts in the scene? For example, X is on Y, X is under Y, X is near Y.

@adrianJW421 commented on GitHub (Feb 19, 2025): > Hi [@adrianJW421](https://github.com/adrianJW421), can you share your prompts in your cases? Thx for relying. Currently in my experiment, I used following prompts: 1) Can you describe the scene in the images, highlighting the locations of obejcts in the scene? For example, X is on Y, X is under Y, X is near Y. This some times leads to immediate 'EOS' I also try to improve with <[grounding]>, it did looks better(I mean, improved a bit). 2) <[grounding]>Can you describe the scene in the images, highlighting the locations of obejcts in the scene? For example, X is on Y, X is under Y, X is near Y.

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepSeek-VL2#75

No description provided.

Rows
Columns