Some questions about subtitles

[Nastu-Ho]

1. How did you get the subtitles for each image frame in the training? Does the training dataset come with it?
2. Are subtitles different for different frames of the same video?

[Nastu-Ho]

Did the linear layer and the LLM participate in training together in all three stages?

[Nastu-Ho]

Whether subtitles are also involved in training in the third stage？

[KerolosAtef]

Hello @Nastu-Ho
For the subtitles in training datasets:
for CMD the subtitles are provided. and for the Video-ChatGPT and Webvid , I used whisper model to get the subtitles.

For each video we sample 45 frames for llama2, and each frame has the subtitles. for instance, frame 2 will have all the subtitles from frame 2 to frame 3.
for (mistral sample 90 frames instead of 45).

yes, the linear layer and the LLM trained for the 3 stages.

Regarding subtitles in the third stage: yes, subtitle is included.