@florentromanet5439
Thanks Matt! I'm really interested in training and/or fine-tuning models with SMBs customer's data on-premise. Literally inside the company on their network mainly for data compliance. Swiss here 🇨🇭😁 Would love to see more of such use cases. Great job again.
@Mavi-m1v
Hi Matt, thanks for your great work here. You make complicated things look so easy. Love that. Could you please show how to train your model with local data? What format and how your local data should be. Thanks, M
@goonie79
Thank you so much for this video, this is exactly what I was looking for since I didn't want to deal with the collab headaches and timeout issues.
@peteraldrich8385
I built my own Jupiter hub session rather than depending on collab. I love the video because it gives me more ideas on what I can implement into my home lab
@escain
One of the main goals for Ollama is (IMHO) the capability to run locally without SAS or data protection issues. This video brings us closer to fine tuning locally, but we still have direct access to huggingface (making us vulnerable to model removed or changed). Also, for me, being bound to NVidia monopolio and it CUDA closed technology is a No-Go. I think that I can make the little remaining road alone to get this working locally and with ROCM / OpenCL. Thanks for this great video. 🇨🇭😁
@vertigoz
Is it me or the sound is a bit off?
@user-df2ow9gz3v
Thanks Matt for that great video.
@tomek-858
Kudos for suepr smooth 'subscribe' animation at 8:30 ;)
@jonathantremblay3469
Okay, a little throwback on this video. I finally made a ton of work using this video as a base. First, I would strongly suggest to use Kaggle instead of Google collab. They have P100 for a stronger fine-tuning you, but also t4 and TPUs. And well you have 30h per week with 12h of runs every shot. That's so cool. The hardest part was to make a decent dataset because I wanted to make mine instead of using an already made up one. And Qloras are maybe a bit more interesting than lora when you're memory limited. All the tweaking, yeah Claude saved my butt many times. Everything went so fine, that was kinda easy. I then converted the notebook to use on windows. My machine is a win10 ltsc WITHOUT wsl2. That was only but a freaking pain... Claude or any coders out there were NO USE. Triton is hard to install... That said, there IS a version that works on windows for the 11.8 and 12.x Nvidia ski toolkit. That took me 3 days to figure out... It's on hugginface...! :) However, took me a while but I finally made it work on my 3090.
@daskasspatzle2396
Thank you for sharing the knowledge. I am still struggling with understanding what a "model" is at all... Although i do not understand all that stuff, i am always amazed how you manage to explain the single steps:finger-red-number-one:
@FleaMarketSocialist
Bro, I know your R1 video is going to be fire as fuck!🔥
@sanjeevgupta8435
Hi Matt, Thanks for educating us on fine tuning and sharing your valuable tips.. can you pls share some references on preparation of datasets for code LLMs. I am working on fine tuning the LLM s to make coding faster .. also another query - is fine tuning a better approach or is RAG better.. thanks once again
@Mede_N
Thanks for the video, arrived shortly after I tried out Unsloth for the first time. The biggest hurdle was - like you said - the data preparation :). Btw. something seems to be wrong with your mic-sound in this video: It sounds clipped and undersampled.
@janielj5963
Hello Matt, You are amazing!. thanks for your content. I have a request., Can you create a video on foundational model, and is it simply fine tuning a specific data
@RetroverseJack
great tutorial Matt! Could you shed some light as far as converting jsonl file to parquet format or just using jsonl itself? Thanks.
@Xuexi4Fun
Hi Matt, Unsloth recently reveals its train R1 reasoning model with Unsloth (GRPO), have you tried it out? Looking forward to seeing your tutorial on it.
@Luisow748
Hey Matt! Thanks for the content! I would suggest you to talk about creating and using a neo4j graph RAG using Ollama.
@anshulsingh8326
Can't we train another model other then the list? I wanted to train deepseek coder v2
@mercedesbenzla911kenya3
Would love to see you train with sample data and wee review the untrained and trained outputs
@mahakleung6992