4 pointsby HappyPablo5 hours ago2 comments
  • teamcubitflow4 hours ago
    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.
    • HappyPablo4 hours ago
      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
  • HappyPablo4 hours ago
    [dead]