I'm surprised that kind of captioning came from a 2B model; glad the fine tuning process actually shows a deliberate approach to making qwen 3.5 into essentially a new model of it's kind.
hey this is shubham, yeah Qwen3.5VL is awesome and it's training vocab is quiet strong so with the right data curation you can prolly take it into a bunch of other narrow tasks eg: we trying to fine-tune it to use SAM3 in a loop for segmentation tasks in the videos