Inquiry about the MLLM and prompts used for recaptioning

[w1oves]

Hi authors,
Thanks for your great work! I am currently studying your dataset and have a few questions regarding the recaptioning process mentioned:

1. Model Selection: Could you clarify which Multimodal Large Language Model (MLLM) was used for the recaptioning pipeline?
2. Prompt Details: What specific prompts were used to guide the model during this process?
3. Experimental Analysis: Does the paper include any experimental analysis or ablation studies specifically focusing on the impact or quality of these recaption results? If so, could you point me to the relevant section?

Thanks in advance for your time and help!

[kaichengyang0828]

Thanks for your attention. As described in the paper, the DanQing dataset itself does not contain recaptions; it is internet data obtained after cleaning. However, internally we have used different models for recaption. The synthetic captions we ultimately open-sourced were generated using GLM4.1V-9B. We have compared the captions generated by this model and Qwen2.5-VL 32B on our internal evaluation set and found their performance to be comparable.