So this is the master class, in practical LLM fine-tuning with the latest cutting edge technology. I have searching for such a comprehensive tutorial for so long. Thanks!
I needed this, ty so much. im going to be buying your plan from your site at some point in the near future, until then i am building a GPT in chatgpt what can fully handle the process of creating high quality synthetic datasets and ill link it here when its complete. please respond to this if youd like to use it so i remember to link it!
Yes, unsloth is single gpu. The current script though allows you to set unsloth=True or False, in the false case you can run transformers, which will allow you to run multi-gpu.
For this vid, the notebooks are only available for those who have purchased lifetime access to the ADVANCED-fine-tuning repo (see trelis.com/advanced-fine-tuning). I'll keep in mind to blend in some videos with some more free content going forward. I'm trying to get a good balance to ensure the business is well funded and growing so I can keep making content.
good Q. It's not that straightforward... - In principle, the longer the responses the more noise will take over - BUT, for shorter answers there are fewer possible answers to choose from and, as such, it is more likely to hit a bad answer, whereas with more tokens the model can work its way towards a better answer. That's my current intuition, definitely not gospel.
@@TrelisResearch When i say longer answer i mean to put the reasoning or justification of answer before final answer like that the relevant informations (so the attentions) are near the final answer
So this is the master class, in practical LLM fine-tuning with the latest cutting edge technology. I have searching for such a comprehensive tutorial for so long. Thanks!
Thanks! You’re welcome
Amazing content! I am learning a lot.
Thanks for this great video
I needed this, ty so much. im going to be buying your plan from your site at some point in the near future, until then i am building a GPT in chatgpt what can fully handle the process of creating high quality synthetic datasets and ill link it here when its complete. please respond to this if youd like to use it so i remember to link it!
Great stuff! Sure
Thanks for sharing! Does unsloth currently not support multi-GPU fine-tuning? The current script is only suitable for single GPU fine-tuning?
Yes, unsloth is single gpu.
The current script though allows you to set unsloth=True or False, in the false case you can run transformers, which will allow you to run multi-gpu.
@@TrelisResearch Get it!
Would you share the colab link?
For this vid, the notebooks are only available for those who have purchased lifetime access to the ADVANCED-fine-tuning repo (see trelis.com/advanced-fine-tuning).
I'll keep in mind to blend in some videos with some more free content going forward. I'm trying to get a good balance to ensure the business is well funded and growing so I can keep making content.
@@TrelisResearch Oh…what a pity:(
Anyway, it’s really a helpful tutorial! Thanks!
The longer the answer is, the better the answer is as it's autoregressive model and attention will be better no ?
good Q. It's not that straightforward...
- In principle, the longer the responses the more noise will take over
- BUT, for shorter answers there are fewer possible answers to choose from and, as such, it is more likely to hit a bad answer, whereas with more tokens the model can work its way towards a better answer.
That's my current intuition, definitely not gospel.
@@TrelisResearch When i say longer answer i mean to put the reasoning or justification of answer before final answer like that the relevant informations (so the attentions) are near the final answer