2 pointsby timxzz6 hours ago3 comments
  • freakynit6 hours ago
    This is so good. Is it kinda similar in concept to QLoRA training?
    • zqiu5 hours ago
      Hi, thank you so much :) Yes, you are very sharp, to make truely it scalable, we have to finetune quantized LLMs directly, similar ideas as QLoRA. The algorithmic innovation is based on our research about orthogonal finetuning OFT and QOFT, which has demonstrated better training stability when finetuning quantized base models in such a scale.
      • freakynit3 hours ago
        Thank you for the information, creating this and releasing the research behind it. Truly appreciate.
  • kendy19926 hours ago
    Looks very cool!!!
    • zqiu5 hours ago
      Thank you so much, glad you liked it :)
  • crookedusage906 hours ago
    wow impressive
    • zqiu5 hours ago
      Thank you so much :)