Thanks for your great work!
I can't reproduce the performance when I finetune LLaVA with LingoQA dataset.

Here please help me to align the experient setting

1. The single image selected is the middle one from the image sequence.
2. During training, is the multi-round QA or single-round QA format adopted?
3. During fine-tuning, as the LLaVA fine-tuning approach is used, freeze the visual encoder weights and fine-tune the weights of the projection layer and the whole LLM in LLaVA.