关于stage2的json文件

[liziming5353]

请问你们stage2做微调的时候是默认使用llava_v1_5_mix665k_with_video_chatgpt_maxtime_5min.json这个文件吗？如果是的话，为什么要把视频限制在5分钟以内呢？

[yanwei-li]

It's the default one for the model with vicuna 13B (the one for 7B without time limit), because we need to avoid the out-of-memory problem during training if the video is too long.

[liziming5353]

请问有预估需要多少内存吗？如果我的内存够用，是不是就不用限制5分钟了呢

[yanwei-li]

Hi, for vicuna 7B, it requires <40G in my experiments, and 80G for vicuna 13B.

[liziming5353]

Thanks! Also, may I ask how much loss can you get in stages 1, 2, and 3? I want to check if my training is correct. Now in stage 1 I get loss=2 and in stage 2 I get loss=1

[yanwei-li]

Hi, the loss for stages 1, 2, and 3 is about 1.9-2.1, 0.8-1.0, 1.3-1.5, respectively.

[liziming5353]

Got it!