mneedham
commented
9 months ago Thinking about this more - probably the best way would be to find some things where LLaVA works well and then use LLaVA as the ground truth. We can then see how close BAKLLaVA gets.
- Do some similarity testing with embeddings
- And then use Ragas to compare
Need to see how long the video gets so we can see whether we do it all in one video
Compare this with LLaVA - https://ollama.ai/library/bakllava