Synthetic Data Generation and Fine tuning (OpenAI GPT4o or Llama 3)

แชร์
ฝัง
  • เผยแพร่เมื่อ 11 ม.ค. 2025

ความคิดเห็น • 15

  • @nirsarkar
    @nirsarkar หลายเดือนก่อน

    So this is the master class, in practical LLM fine-tuning with the latest cutting edge technology. I have searching for such a comprehensive tutorial for so long. Thanks!

    • @TrelisResearch
      @TrelisResearch  หลายเดือนก่อน

      Thanks! You’re welcome

  • @EminTemiz
    @EminTemiz 3 หลายเดือนก่อน +1

    Amazing content! I am learning a lot.

  • @loicbaconnier9150
    @loicbaconnier9150 3 หลายเดือนก่อน

    Thanks for this great video

  • @BirdManPhil
    @BirdManPhil 3 หลายเดือนก่อน +1

    I needed this, ty so much. im going to be buying your plan from your site at some point in the near future, until then i am building a GPT in chatgpt what can fully handle the process of creating high quality synthetic datasets and ill link it here when its complete. please respond to this if youd like to use it so i remember to link it!

  • @remixa-jp6rj
    @remixa-jp6rj 3 หลายเดือนก่อน

    Thanks for sharing! Does unsloth currently not support multi-GPU fine-tuning? The current script is only suitable for single GPU fine-tuning?

    • @TrelisResearch
      @TrelisResearch  3 หลายเดือนก่อน +1

      Yes, unsloth is single gpu.
      The current script though allows you to set unsloth=True or False, in the false case you can run transformers, which will allow you to run multi-gpu.

    • @remixa-jp6rj
      @remixa-jp6rj 3 หลายเดือนก่อน

      @@TrelisResearch Get it!

  • @minyi3203
    @minyi3203 3 หลายเดือนก่อน

    Would you share the colab link?

    • @TrelisResearch
      @TrelisResearch  3 หลายเดือนก่อน

      For this vid, the notebooks are only available for those who have purchased lifetime access to the ADVANCED-fine-tuning repo (see trelis.com/advanced-fine-tuning).
      I'll keep in mind to blend in some videos with some more free content going forward. I'm trying to get a good balance to ensure the business is well funded and growing so I can keep making content.

    • @minyi3203
      @minyi3203 3 หลายเดือนก่อน +1

      @@TrelisResearch Oh…what a pity:(
      Anyway, it’s really a helpful tutorial! Thanks!

  • @loicbaconnier9150
    @loicbaconnier9150 3 หลายเดือนก่อน

    The longer the answer is, the better the answer is as it's autoregressive model and attention will be better no ?

    • @TrelisResearch
      @TrelisResearch  3 หลายเดือนก่อน +1

      good Q. It's not that straightforward...
      - In principle, the longer the responses the more noise will take over
      - BUT, for shorter answers there are fewer possible answers to choose from and, as such, it is more likely to hit a bad answer, whereas with more tokens the model can work its way towards a better answer.
      That's my current intuition, definitely not gospel.

    • @loicbaconnier9150
      @loicbaconnier9150 3 หลายเดือนก่อน

      @@TrelisResearch When i say longer answer i mean to put the reasoning or justification of answer before final answer like that the relevant informations (so the attentions) are near the final answer