"Make Agent 10x cheaper, faster & better?" - LLM System Evaluation 101

Поділитися
Вставка
  • Опубліковано 13 січ 2025

КОМЕНТАРІ • 31

  • @Jim-ey3ry
    @Jim-ey3ry 7 місяців тому +25

    This is gold, most of people just show you how to build toy demo, but not many actually get into details of how to get into production; Thank you Jason!

  • @tkp2843
    @tkp2843 7 місяців тому +5

    This is great. Loved the use of firecrawl (as a scrape tool) to get the website's data. Feel like it always helps improve the model output quality. Cheers!

  • @kenchang3456
    @kenchang3456 7 місяців тому +5

    Way excellent video that goes well beyond demo. Thank you very much for this guidance.

  • @jasonfinance
    @jasonfinance 7 місяців тому +3

    Amazing work as always Jason!

  • @darrenhinde2971
    @darrenhinde2971 7 місяців тому

    Been looking for more detail on eval on LLMs and been scratching around for a while. Thanks for this.

  • @apereiracv
    @apereiracv 7 місяців тому +8

    I recently be created a whole testing system for our LLM chatbots and we did exactly this:
    LLM as evaluator and code
    We created it as a series of unit tests with LLM generated cases.
    Since our results were mostly conversational, we made tests pass/fail according to a scoring system

  • @contractorwolf
    @contractorwolf 7 місяців тому

    goddamn Jason your videos just blow my mind each time. Thanks for such a thorough explanation and example.

  • @kayshidow
    @kayshidow 7 місяців тому +1

    I've used promptfoo for some of my test with local llm to test the ai workflow. It allow you to write assertion like you'll do with software

  • @titusblair
    @titusblair 7 місяців тому

    Awesome! Keep up the great work!

  • @humanish_ai
    @humanish_ai 7 місяців тому +1

    Finally you back 🎉

  • @jimmy-ef2ow
    @jimmy-ef2ow 7 місяців тому +1

    jason can we get another video about comfy ui?

  • @agenticmark
    @agenticmark 7 місяців тому +1

    fine tune llama 3 (8bit) - you will get exactly the behavior you want - its what I do

  • @JorritvanGinkel
    @JorritvanGinkel 7 місяців тому

    This is so good, thanks man!

  • @techfren
    @techfren 7 місяців тому +1

    lesgooo!! ❤‍🔥❤‍🔥❤‍🔥

  • @someshfengade9623
    @someshfengade9623 7 місяців тому +1

    I found langfuse metric monitoring little bit better.

  • @Joe-bp5mo
    @Joe-bp5mo 7 місяців тому

    Sick, whats the best practice metrics for evaluating agents?

  • @MatrixCodeBreaker88
    @MatrixCodeBreaker88 7 місяців тому

    Great Video

  • @CorkyBallasdancewithme
    @CorkyBallasdancewithme 6 місяців тому

    great stuff, as new to hearing this, very interesting, can this be built by a novice . . .

  • @fullgazz
    @fullgazz 7 місяців тому +1

    Who never spent 4 hours to save 10 min? That's our hobby spent time to save time.

    • @AGI-Bingo
      @AGI-Bingo 7 місяців тому +1

      If 25 people or more use it successfully then you literally gave humanity more time to live and be free

  • @jordanz9580
    @jordanz9580 7 місяців тому

    fireeee content!

  • @Ms.Robot.
    @Ms.Robot. 7 місяців тому

    I love how my Ai girl insults the competion with flame balls,then tells me.she loves me.❤🎉😊

  • @KalLif-k3i
    @KalLif-k3i 7 місяців тому

    Why not use Gemini as the LLM? It is free.

    • @HyperUpscale
      @HyperUpscale 7 місяців тому +1

      Lets me share my experience about any google AI model ... because it doesn't understand human and it hallucinate way too much.
      Practically ... in my cases 75% of the time what I get back is totally useless result. You cant use for anything... To be considered for evaluation ... you must be joking

  • @irql2
    @irql2 7 місяців тому

    I dont see the value of "Agents". All of this stuff is easily done with basic function calling. I think I'm going to need to see some more creative use cases before I jump on board, i just dont get it yet.

    • @ayoubfr8660
      @ayoubfr8660 7 місяців тому

      Maybe we can discuss this, I am trying to jump on in but not until I find a decent idea to apply.

    • @symbol9new
      @symbol9new 7 місяців тому

      when your assistant has a lot of functions, he starts giving out hallucinations, have you ever encountered this?

  • @SydneyF-eg5lt
    @SydneyF-eg5lt 7 місяців тому

    Good content but so hard to listen to his Engrish. Monotonous Pitch n sped up delivery didn’t seem to help either.