Your Computer, Your Models, Your Rules - Transformer Lab
ฝัง
- เผยแพร่เมื่อ 9 ก.พ. 2025
- Ali and Tony demonstrate the power of Transformer Lab, an open-source platform that allows anyone to build, tune, and run LLMs locally, without writing code.
Learn more: transformerlab.ai
Get the code: github.com/tra...
Mozilla Builders helps developers create transformative open-source AI projects through collaborations, programming, and community. Check out the projects we’re supporting: builders.mozil...
2:23 - Start here (skip intro)
Looks like I stumbled on a treasure trove. Very nice thanks you for that
Excellent , you did open source a very good service thank you for the effort!
OPEN-SOURCE RULES!!!🎉 Thank you TransformerLabs!❤
This team is on 🔥🔥! Keep it up
This looks fantastic!
You said you were fine tuning the base model, but you somehow magically ended up with a fine tuned instruct model. Are you sure you aren’t fine tuning the instruct model?
This is awesome, I will be testing it out
great !!! i will try out this weekend 🙂
Exactly what i need. thanks
Love it! Great job
Wow this is great!
stopped with errors, while install torch==2.5.1, macos 12.7.6
not for me :(
Create an issue at the repo, it is open source
Hi, I installed it on my Windows 11 PC, it says that I need to install "WSL", I search that, and the Microsoft page says that I have to run the command wsl --install, I did that and seems to be done ok. But when opening the app is still saying that I need WSL (I uninstalled and installed the app again and have the same issue). The app version that I installed is 0.9.3.
Any help on how I can get the app running? Thanks!
same, wsl is installed and active
Did you try rebooting
@@917439😂
So true, rebooting solves 95% of the issues
@@917439 yes
you probably need to add a linux distribution to your wsl, ubuntu 24.04 LTS should be good and can be directly downloaded from the microsoft store, reboot after that to make sure it's installed properly
This is interesting, however I've been told that fine-tuning modern models can be a delicate process that might have to be customized for each model. How is this achieved?
Ironically I just tried transformer lab and the first issue I hit was python package and versioning issues.
"You've probable seen a wall of python errors like this one" every single day, buddy
I want a way to train a small model that is already good in like language translation, math or programming and improve them on that field for example give it a whole new documentation on x framework, I could just scrape it ask ai to generate the test rewards etc, the data to be trained on based on the documents provided and then just go, test itself and self improve
Thanks
Nice!
Sooooooo cool!
guess again going on the odd direction AI is eating UI but it's cool like for now it's great really
Write it in c++
Write in Rust, you cowards!
@@TheMaxAwesome never
Write it in anything that is not python, I'd love C, C++ or Rust :D
woow cool
cool
1B non-specialized are not worth massively deploying yet 😕
But they are getting better, its nice to see the tooling get developed and it will be ready when small LMs are ready.
Most people NEED specialized models with a generalized base. They don’t actually need a fully generic model, just enough to get a good set of navigational language around a particular domain.
Tools like this or IBM’s InstructLab are great for models that companies want to use in-house.
Mac users don’t know how to full screen an app so now I’m having to squint. Smh
Qwen2? How old is your presentation?
So, the only way an LLM can learn a subject you want to teach it is Q&A? Seems tedious and unnecessary. Why not just give it a set of PDFs, Webpages, Official Documentation, Rule guides, Governmental references etc. If you have to teach it by giving it the answer to every single possible query and permutation of that query, you will be training for a lifetime on something as simple as a sports game.
Yes exactly my question
because the raw data within PDFs, webpages, docs, etc fundementally has no meaning, you need to understand that training an LLM is like trying to teach an alien baby about humans, you NEED to provide a clear concrete relationship which is a q -> a relationship.
if you were to train an llm just on raw data without any relationships then it will only memorize what's written on it and would know nothing about what any of it means
LLMs don't need to be trained on questions and answers. However, because LLMs work by predicting the next token (fragment of a word) then when an LLM that has not been trained on questions and answers is asked a question by a human the tokens that an LLM might respond with might not be what we consider an answer, much less what we could consider a good answer.
So the best way we've found to increase the likelihood that an LLM answers with a good answer is to train it on questions and good answers.
Teaching on QnA is supervised fine tuning. Teaching next word prediction on pdfs is pre training. There is a difference for a reason. SFT is needed so that LLM can answer in QnA format.
If you don't want your model to just output random blurbs of text your need to train it on completely intact and grammatically correct samples, otherwise it could just stop talking in the middle of a
Yeah, Python sucks. That's why I switched to Rust, C, and R.
4:10 as per usual, local llms suck.
Bro fix your browser first
Firefox best
@HateAcceptedButThinkAgain lol nice joke
As always, you are focusing on the wrong problem. You should have spoken with a domain expert first to gain a proper understanding. I hope my donation didn’t go toward funding you two.
Maybe, interested to know what the right problem is?
So please do elaborate
All snark and no substance, try to give us something useful instead of generic LLM snobbery.