19 Tips to Better AI Fine Tuning

Поділитися
Вставка
  • Опубліковано 24 січ 2025

КОМЕНТАРІ • 82

  • @michalvis2476
    @michalvis2476 15 днів тому +26

    Let's have an MLX!

  • @johnpummill8903
    @johnpummill8903 11 днів тому +3

    Great tutorial. Really appreciate your teaching style. Unsloth is what I am researching right now.

  • @mecha-intel
    @mecha-intel 15 днів тому +17

    Thank you for the excellent video. Exactly what I need. Can you please start with MLX?

  • @RalphKrausse
    @RalphKrausse 14 днів тому +3

    Like always, awesome stuff. I vote for MLX but do want to see all of them. Excited and thanks for producing these.

  • @jrnmadsen2710
    @jrnmadsen2710 15 днів тому +8

    Great !!! Macbook Silicon user here. Can't wait (but not holding my breath) :)

  • @CGTISOLUCOESCLOUD-JosemarMigow
    @CGTISOLUCOESCLOUD-JosemarMigow 13 днів тому +2

    Great introduction!
    I would like to see unsloth first

  • @kevinmcquown
    @kevinmcquown 15 днів тому +9

    definitely interested in MLX since I'm all apple silicon

  • @JanKrummrey
    @JanKrummrey 15 днів тому +5

    I'd love to see more information on MLX and maybe you could cover also using Ollama with languages other than English.

  • @oliverbarreto1838
    @oliverbarreto1838 13 днів тому +1

    Great video again Matt !!! Knowing the difference to not make mistake on when to choose Fine Tunnng over RAG really deserves a long video like this one.
    I prefer MLX since I have a mac
    Also importan… please plan a video to know different hardware setups to train, and to use to use local models for inference on a home lab environment, how to use very large models like llama 3.3 70B in a homelab, pros and cons
    Thanks again

  • @moviechristian
    @moviechristian 14 днів тому +4

    Unsloth tutorial for start, then the others
    Thanks for share your knowledge

  • @mahdihosseini1085
    @mahdihosseini1085 14 днів тому

    Well done!
    Really great tips for fine tuning
    I'm looking forward to the next episodes

  • @RasmusRasmussen
    @RasmusRasmussen 15 днів тому +1

    Another great video full of info, I wish I’d known when I first started messing with AI models. Thank you!

  • @tarmiziizzuddin337
    @tarmiziizzuddin337 15 днів тому

    Clear and concise introduction on fine tuning LLM, great video as always!

  • @DanielAI-p5x
    @DanielAI-p5x 2 дні тому

    Excellent tutorial! Thanks Matt!

  • @mattchinnock7064
    @mattchinnock7064 14 днів тому

    Such an excellent channel. Always appreciate your content!

  • @قيمنقعبود-ب2ل
    @قيمنقعبود-ب2ل 11 днів тому

    Nice video. One of the best in this field

  • @Patrick-rd9qo
    @Patrick-rd9qo 12 днів тому

    Love your use of analogies!

  • @Alex-os5co
    @Alex-os5co 15 днів тому

    Awesome, soo keen to follow along! Thanks Matt. I’m going to try them all but wouldn’t mind starting with MLX first

  • @Armenian-abrank
    @Armenian-abrank 12 днів тому

    Fantastic explanation! 😊

  • @mahdihosseini1085
    @mahdihosseini1085 14 днів тому

    As a case study I want to learn a little about training a model from scratch!
    I hope to see it soon!

  • @PhilEhI
    @PhilEhI 15 днів тому +2

    Thanks. I was misunderstanding fine tuning text models. But what about fine tuning a text and vision model. Could you tune it to recognize images it wasnt trained on?

  • @johnnygri99
    @johnnygri99 14 днів тому

    Thanks Matt for this great introduction. As an M1 Max user I'll be looking forward to learning using that.

  • @drp111
    @drp111 15 днів тому +2

    Interesting! I wonder if its possible to train my long system prompt into the model itself.

  • @andrearomito3335
    @andrearomito3335 14 днів тому

    Grwat video! I was really looking forward to you explaining finetuning in regards to Ollama! I'd vot either one method, as long as it's done all locally (no online service, and oviously Lora and Qlora oriented) and can be done on windows!

  • @ISK_VAGR
    @ISK_VAGR 15 днів тому +2

    @Matt as always super useful. I am down for the MLX fine-tuning first. I became a member from your channel today. Keep doing this high quality content and helping us to navigate Ollama and Llama models potential. 🦾🦾🦾

  • @juanantonionavarrojimenez2966
    @juanantonionavarrojimenez2966 15 днів тому +1

    This video is gold.

  • @raptornc1275
    @raptornc1275 15 днів тому +2

    Great Video!! I would love to see a video on training with TextGen WebUI. It supports Lora and QLora and with the Training_Pro ext. it has dozens of setting including some experimental ones, it also has a built in training//loss graph so you can track your progress. best of all it supports Multple GPU's so you can train bigger models locally. The problem is it lacks good documentation so I dont know what half of the options do. I would be cool to see you cover them.

  • @aibeginnertutorials
    @aibeginnertutorials 15 днів тому

    Excellent as always!!! Thanks again

  • @roygatz
    @roygatz 15 днів тому

    can't wait, you are the best!

  • @miladnekooei8468
    @miladnekooei8468 8 днів тому

    Great!
    maybe we need playlist to cover all details about fine tunning...

  • @BORCHLEO
    @BORCHLEO 15 днів тому +1

    I know its not on the list but I would also love to see how to finetune with Autotrainer, Trl, and Peft with HF :) thank you for this amazing video!

    • @BORCHLEO
      @BORCHLEO 15 днів тому +2

      i am SUPER excited for axolotyl tho!

    • @AdamBrusselback
      @AdamBrusselback 15 днів тому +2

      @@BORCHLEO keep in mind that axolotl wraps TRL and PEFT, so the same info should apply to both..

    • @BORCHLEO
      @BORCHLEO 11 днів тому

      @@AdamBrusselback sweet!

  • @biagiomaf5951
    @biagiomaf5951 14 днів тому +2

    Hello great teacher, unsloth would be best for me. Thanks

  • @BadBite
    @BadBite 15 днів тому

    You are a great instructor!

  • @rogueturnip
    @rogueturnip 15 днів тому +3

    If you can start with MLX that would be amazing!

  • @Tripp111
    @Tripp111 15 днів тому +2

    Thank you. Unsloth, please.

  • @tylerlindsay9263
    @tylerlindsay9263 14 днів тому

    I am interested in creating better training data. Perhaps explore how generating synthetic training data works.

  • @VinceThyng
    @VinceThyng 11 днів тому

    Thanks, I found this video helpful. I am interested in tuning a vision model usable through ollama for Lego identification. I am not sure any of the models actually know the Lego part numbers, so I suspect I need to train a model also. One goal is tell me what this Lego part numbers is.

  • @TheRob2079
    @TheRob2079 15 днів тому

    Great videos I like your content. Can you review fine tuning an image model? I am thinking llama-3.2-11B the use case I have been considering is to fine tune it to better recognize a person's hand writing style to more accurately convert written documents to text as part of a pipeline to then have another fine tuned model read those text notes and write a short report on them. Just a thought. Thanks!

  • @zuhepix
    @zuhepix 14 днів тому

    Nice didactic video. Would love to see mlx first

  • @DigitalDesignET
    @DigitalDesignET 13 днів тому

    You're the best! BTW MLX please.

  • @NLPprompter
    @NLPprompter 15 днів тому

    is there any cloud free fine tune services guys?

  • @pjain001
    @pjain001 14 днів тому

    Fine tuning and RAG with MLX, please.

  • @steenharsted
    @steenharsted 15 днів тому +4

    Unsloth please

  • @invalidred
    @invalidred 11 днів тому +1

    Unsloth or mlx

  • @مقاطعمترجمة-ش8ث
    @مقاطعمترجمة-ش8ث День тому

    [lease do more videos about this

  • @bio_code
    @bio_code 15 днів тому +4

    Please turn off that UA-cam translation thing. On German it’s horrible. But thanks for the video

  • @raiden7378
    @raiden7378 15 днів тому

    Great content. The sound quality is pretty bad. Thought I’d let you for next time

    • @technovangelist
      @technovangelist  15 днів тому

      tell me more about your setup. Would love to know what your setup is like where it sounds bad. Tested this one on lots of devices and sounds great.

    • @raiden7378
      @raiden7378 14 днів тому

      @ iPhone 16 pro max. Pretty sure it’s not my device. I tested it with multiple videos to make sure that it’s not my phone right before posting the comment above

    • @technovangelist
      @technovangelist  14 днів тому

      Sorry to pester but can you define bad? I can’t hear anything wrong

    • @raiden7378
      @raiden7378 13 днів тому

      @ no worries. It’s probably just on my end.

  • @-........
    @-........ 15 днів тому

    Da Capo

  •  14 днів тому

    MLX 🥰

  • @NarkeEmpire
    @NarkeEmpire 13 днів тому

    So... What if i need to teach him new information? 🥶

  • @mbarsot
    @mbarsot 15 днів тому +3

    Mlx

  • @paulmiller591
    @paulmiller591 15 днів тому +3

    MLX

  • @snygg-johan9958
    @snygg-johan9958 15 днів тому +1

    @technovangelist Hey man, love your content. Please dont put race cars and cool images as thumbnails, its better if we can see a picture of you, so we quickly can spot that its you who put out a new video and its time to tune in.
    With ai generated thumbnails, you kind of blend in with all other channels.
    My notifications are already overloaded with channels, making it near unusable, so I rely on the youtube algorithm to show me videos Im interested in based on how much I watch the channels.
    ...and regarding fine tuning mlx, unsloth, and finally axolotl :-)

    • @technovangelist
      @technovangelist  15 днів тому +2

      thanks. actually every thumbnail i have made in nearly 2 years has used AI generated imagery that I then tweak. But I updated it to use my face. Then ran the A/B test with the new one and old one. Within an hour it was saying 65% of views were coming from the thumb with my pic in it.

    • @snygg-johan9958
      @snygg-johan9958 15 днів тому

      @@technovangelist Oh :-) Well, Im no content creator, so I guess I should leave it up to you since you have the data of what works. In any case, you have really good stuff on your channel :-)

    • @technovangelist
      @technovangelist  15 днів тому

      thanks for mentioning it. its made a great difference on the view counts

  • @simplir5644
    @simplir5644 14 днів тому

    Mlx first 🎉

  • @Cyberspider76
    @Cyberspider76 14 днів тому

    MLX please!!!

  • @yangp2897
    @yangp2897 15 днів тому

    MLX please

  • @growdelan9855
    @growdelan9855 14 днів тому

    first MLX or riot!

  • @greenbes66
    @greenbes66 15 днів тому

    MLX, please

  • @NABZ028
    @NABZ028 15 днів тому

    MLX, please but looks like you already covered it in ua-cam.com/video/3UQ7GY9hNwk/v-deo.html ? :) Anyway, thank you for this excellent video, you have very well explained what fine-tuning is and the difference with RAG.

  • @BigAura
    @BigAura 15 днів тому

    MLX, the others have already been covered.

    • @technovangelist
      @technovangelist  15 днів тому

      actually, MLX is the only one I have covered, but I plan to do it again with updates. And none of the coverage on the other tools have been that good I think. But MLX seems to be the winner so far

    • @BigAura
      @BigAura 14 днів тому

      Others covered by other youtubers. MLX not much. Probably a good niche.@@technovangelist

  • @chunwaichan6807
    @chunwaichan6807 12 днів тому

    let’s have MLX first.

  • @elgimblett
    @elgimblett 15 днів тому +1

    Axolotl get my vote for the first method . Thanks Matt

  • @zandanshah
    @zandanshah 14 днів тому

    Axolotl Open-Source tool, I will vote for this.

  • @sebastiangrosso5405
    @sebastiangrosso5405 7 днів тому

    Ai Spanish 🙏

    • @technovangelist
      @technovangelist  6 днів тому

      apparently the ai subs and dubs are terrible, so they are off

  • @MatteDeiMatte
    @MatteDeiMatte 14 днів тому

    Axolotl AI (smash the like button to this comment)

  • @eduardoribeiro6267
    @eduardoribeiro6267 13 днів тому +2

    Unsloth please

  • @steen_rabol
    @steen_rabol 7 днів тому

    MLX, please