BytedanceDouyinContent

/

SAIL-VL-1d5-2B

like 12

Follow

BytedanceDouyinContent 83

I would recommend the toolkit from InternVL: https://github.com/OpenGVLab/InternVL/tree/main, because our model shares a similar architecture with InternVL 2~3, it is very easy to use SAIL architecture in InternVL code, please refer to the repo for more details.

Sir i have already tried many times and got so many errors, would you be a little more specific? i will be grateful to you . please help me .
The more you elaborate, is better to understand as your great work will be used for the betterment of mankind. Thanks

? Waiting for reply, i loaded the weight successfully, but during training,g loss and norm are 0,1, which doesn't lead to the gradient train, si am struggling with the right script for running it, please someone help lme

Sorry for the late replay, we current can not provide a full script for model training, you might refer to https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/shell/internvl2.5/2nd_finetune/internvl2_5_1b_dynamic_res_2nd_finetune_full.sh for reference.
The error you met seems related to loss computation, it might cause by setting the wrong ground truth, e.g., GT is filled with ignore mark (-100).

Actually, i don't want to know the whole training script or your research, dear, i just want to fine-tune for my data and understand the results.

This is happening, please reply to me on time, kindly, or email me at asifraza151@qq.com

Sorry for the late replay, we current can not provide a full script for model training, you might refer to https://github.com/OpenGVLab/InternVL/blob/main/internvl_chat/shell/internvl2.5/2nd_finetune/internvl2_5_1b_dynamic_res_2nd_finetune_full.sh for reference.
The error you met seems related to loss computation, it might cause by setting the wrong ground truth, e.g., GT is filled with ignore mark (-100).

上面同样的错误，换成原始的internvl3-1b就可以正常微调，得改些代码才可以