Qingyun

Results 78 comments of Qingyun

> [@russellb](https://github.com/russellb) I cannot upgrade to the glibc version, so it cannot work The glibc in your enviroment have to be newer. The root/sudo is also not available to me....

@russellb It can also be resolved by the release of xgrammar, if there will be wheels that will be build with glibc

> should probably open an issue against xgrammar. It doesn't sound like this is something vllm can address. yes, i think so. or is this requirement not too essential? May...

I get a lower performance (platform, free, not locally) ``` Mean Scores Per Category: bleu_score f1_score llm_score count category 1 0.1448 0.2284 0.3582 282 2 0.3471 0.4163 0.4766 321 3...

> > 我得到的表现较低(平台,免费,不是本地) > > ``` > > Mean Scores Per Category: > > bleu_score f1_score llm_score count > > category > > 1 0.1448 0.2284 0.3582 282 > >...

不能替代YOLO啥的是肯定的哇 我们自己微调grounding是没问题的,但是微调检测比较难 如果想要算coco/dota等数据集的mAP更推荐去看看florence-2系列 另外grounding有一个坑,qwen2.5vl的grounding是绝对坐标,数据集组织有一定难度才能work

有难度而已,硬训_能训work的 我自己微调过florence2做密集目标检测都能work效果和fcos差不多 而且怎么可能代替yolo,yolo啥速度,语言模型一个一个推理啥速度

> > 不能替代YOLO啥的是肯定的哇 我们自己微调grounding是没问题的,但是微调检测比较难 如果想要算coco/dota等数据集的mAP更推荐去看看florence-2系列 > > 另外grounding有一个坑,qwen2.5vl的grounding是绝对坐标,数据集组织有一定难度才能work > > 你好呀~ 请我你有没有试过florence-2和Qwen 2vl的对比呀~ 试过,florence-2很容易就能训出来一个检测能跑mAP的模型,qwen2vl如果做dense detection的话不太好训,一直没work 但是grounding上讲(REC),qwen更强,qwen能一开始就有很好的zeroshot grounding能力了

> 应该是的,florence-2有一个dense detection的task。 Qwen系列我看好像是没有的,只能grounding. 感谢分享经验~ 是的,我试过硬训,效果确实不太好,以后有机会还会试一下的,work了来分享