how to train on unstructured data such as GitHub repo code or something like that ? what i mean is i don't have dataset in instruction and answer format but raw text. Do i need to have compulsorily have data in question answer format ?
Will it always give points now for an answer ? Line 1) 2) 3) etc? Or just for the business plan question and whatever you had in your fine tuning dataset? Your reply is greatly appreciated 😊
I'm getting the error: raise KeyError(f"Cache only has {len(self)} layers, attempted to access layer with index {layer_idx}") KeyError: 'Cache only has 0 layers, attempted to access layer with index 0'
Unsloth definitely will be super fast on a 3090 - 2-5x faster and 70% less VRAM usage. Sadly we don't support Macs as of yet :( MLX is probably what you're looking for!
Maybe I do not get something but what is the sense to "train model" if u have sets question-answer? You can just build trivial DB that will serve even better 😅
So given a set of questions and answers, you can use your data to finetune a base LLM like Gemma, Llama or Mistral to make it able to answer new unseen questions. A trivial DB is good, but on questions outside of that DB, a finetuned LLM is able to answer them
Clear and to the point. Thanks!
Thanks for sharing Unsloth and fabulous work on the video! Keep up the great work!
Thank you
how to train on unstructured data such as GitHub repo code or something like that ? what i mean is i don't have dataset in instruction and answer format but raw text. Do i need to have compulsorily have data in question answer format ?
And do you have a tutorial to build the training dataset easily ?
Will it always give points now for an answer ? Line 1) 2) 3) etc? Or just for the business plan question and whatever you had in your fine tuning dataset?
Your reply is greatly appreciated 😊
I'm getting the error:
raise KeyError(f"Cache only has {len(self)} layers, attempted to access layer with index {layer_idx}")
KeyError: 'Cache only has 0 layers, attempted to access layer with index 0'
How do you make and enter a custom dataset? Is there a template?
how do I get conda up and running on wsl?
It says its missing the Triton package but then nor pip or conda can find it. Any solution?
ValueError: Pointer argument (at 2) cannot be accessed from Triton (cpu tensor?)
function calling with small llm
I have obvious questions ... how long does it take ona windows with 3090 ... how long on m1? and what kind of results
Unsloth definitely will be super fast on a 3090 - 2-5x faster and 70% less VRAM usage. Sadly we don't support Macs as of yet :( MLX is probably what you're looking for!
what was the cost for this fine tuning? What can we expect for our use case?
Maybe I do not get something but what is the sense to "train model" if u have sets question-answer?
You can just build trivial DB that will serve even better 😅
So given a set of questions and answers, you can use your data to finetune a base LLM like Gemma, Llama or Mistral to make it able to answer new unseen questions. A trivial DB is good, but on questions outside of that DB, a finetuned LLM is able to answer them
How do we install unsloth ?
How to install Python ? I need to do fine tune