预训练模型与微调相关问题

[cinxiang]

作者您好，
我在复现您的工作时遇到了问题，我不想从头训练一个HicoNet模型，而是直接使用您开源的https://huggingface.co/qihoo360/HiCo_T2I/tree/main模型，但是在我的数据集上不能很好的工作。我想咨询下如何使用lora微调您的HiCo模型，就像您论文中提到的那样。期待您的回复。
另外再咨询下您开源的模型是用哪个数据集训练的？是GRIT还是COCO？

[cbstyle]

1. The open-source HiCo model uses a subset of GRIT20M with a size of 1.2M;please refer to the article for details.
2. Can you provide a detailed introduction to your fine-tuning method? For example, is it the LoRA parameter added to the HiCo collateral? What is the rank setting? What is LR? What does fine-tuning data look like?