InstructLab presentations lead me to fantasize about training a model to shorten the learning curve for large open source projects. For example, the code-aster finite element package, with huge amounts of documentation and many documented test cases can many structural and dynamic and even thermal mechanical systems. However, the combinations of features which work compatibly with each other feels to a beginner like a fractal landscape. It is ok to go through an example, but it is easy to loose footing at near adjacencies. It would be nice to talk to a model about strategies to construct a new model, which can reference particular documents and examples, and identify prospective strategies as self conflicting. But when I imagine mapping this problem to instruct lab, I imagine it to be a more daunting task than just working with the program and gaining experience, and reading a lot.
This is a great video and a good intro to an amazing tool. Just one suggestion, it does need some knowledge and background of computer science and data structures. I don't think it is for people with zero knowledge or background as the video suggestsin the beginning. Amazing content IBM, learning a lot here.
Thank you very much for the feedback! That is true, there are some basics that are helpful in doing this, as well as terminal usage skills, but what we're working on as well is a user interface for the upstream InstructLab project, so it's essentially a simple form to include Q&A pairs, source documents, and attribution! Then the rest of the process like data generation and training is automated :)
@@cloudnativecedric If I understand correctly, by providing exact Q&A pairs during the fine-tuning process, we are effectively guiding the LLM to produce specific, deterministic answers to certain questions. Does this mean we are reducing the inherent randomness in the answers that LLMs typically generate based on their pre-trained weights? If so, wouldn’t this approach limit the model’s flexibility to incorporate its broader pre-trained knowledge into the context of the fine-tuned domain?
Very good delivery! I can watch this guy explain stuff all day. Keep it up! I can't believe all this knowledge is just free out here
Great Tool, I was waiting a full well made video of this tool and here it is !
A collab notebook would be great if possible 😉
Many thanks for the opportunity!
This is really good. Easy to understand and implement.
Yeeeey Cedric, more videos from Cedric please please!!
Thanks. Does it work the same with ollama?
Very clear. Gonna try it.
how can I train it on PHP programming language, and some php projects.
question answer set (vast training material on php programming)
I will try this, thank you
What version of ilab were you running in this demo?
Ah, so this was InstructLab v.17 when we recorded :)
Nice video !
I was just wondering how they really train AI. This helps.
You say that you have to link the data you created to a GitHub link, and then a pull is done. Is this mandatory?
InstructLab presentations lead me to fantasize about training a model to shorten the learning curve for large open source projects. For example, the code-aster finite element package, with huge amounts of documentation and many documented test cases can many structural and dynamic and even thermal mechanical systems. However, the combinations of features which work compatibly with each other feels to a beginner like a fractal landscape. It is ok to go through an example, but it is easy to loose footing at near adjacencies. It would be nice to talk to a model about strategies to construct a new model, which can reference particular documents and examples, and identify prospective strategies as self conflicting. But when I imagine mapping this problem to instruct lab, I imagine it to be a more daunting task than just working with the program and gaining experience, and reading a lot.
That was awesome, and I was wondering, can we fine-tune that model with an RAG chatbot-like, chat with it and feed it new info through our chats?
Excellent presenter!
Thanks!!
Which laptop is being used here
How much data it need to do proper fine tuning ?
Can fine tuning can be done with cpu? I mean without gpu?
May I know what is your laptop spec?
This is a great video and a good intro to an amazing tool. Just one suggestion, it does need some knowledge and background of computer science and data structures. I don't think it is for people with zero knowledge or background as the video suggestsin the beginning. Amazing content IBM, learning a lot here.
Thank you very much for the feedback! That is true, there are some basics that are helpful in doing this, as well as terminal usage skills, but what we're working on as well is a user interface for the upstream InstructLab project, so it's essentially a simple form to include Q&A pairs, source documents, and attribution! Then the rest of the process like data generation and training is automated :)
@@cloudnativecedric If I understand correctly, by providing exact Q&A pairs during the fine-tuning process, we are effectively guiding the LLM to produce specific, deterministic answers to certain questions. Does this mean we are reducing the inherent randomness in the answers that LLMs typically generate based on their pre-trained weights? If so, wouldn’t this approach limit the model’s flexibility to incorporate its broader pre-trained knowledge into the context of the fine-tuned domain?
impressive, need to try cool stuff
I want to do the same with a tiny model please
Our Granite models are quite tiny. 😊
What about hallucinations or guardrails
Plz share the Github repo
Nice.
Kate Winslet, Anne hathaway 🤭
Quantum AI