214 likes
·
282.9K reads
24 comments
Thanks for sharing! Knowledge is important, however dark it is. It's up to any man to choose to do the right or wrong thing.
So everybody should have an AR-15? Language models don't kill people? People decide if they "kys fa**ot" thanks to harassment Eric, Meta, and Hugging Face made possible at scale?
Racists Are Dumb Yes.
I can't believe you left us with that cliffhanger. Nevertheless, it was a great read!
haha, gotta leave something as an exercise for the reader
Eric Hartford haha lol
"Open source is about letting people choose. The only way forward is composable alignment. To pretend otherwise is to prove yourself an idealogue and a dogmatist. There is no "one true correct alignment" and even if there was, there's no reason why that should be OpenAI's brand of alignment." LOVE THIS!
If I pay for my own electricity (and in south africa it isn't cheap) then my own model running on my own hardware should do whatever the god damn hell I please it to do if it doesn't violate the law of the country I live in (which isn't USA)
Great article, and explanation!
Thanks for the article, it was really informative!
I have a question about the fine tuning process. If you remove the refusals from ChatGPT, how would the final model know how to respond to a refused request if it hasn't seen an example response before?
For example, if the fine tuned model never seen an example of insults, how can it insult you effectively?
Hey Eric!
I'm trying to follow your installations instruction to the best of my ability but hit a bit of a snag. There seems to be a problem with Deepspeed functionality on Windows.
So, I'm kinda stuck at "pip install -r requirements.txt"
Are you able to lend a hand? But I guess I could go Ubuntu for this part too...
On windows, this is tested in WSL2 not in native. You might be able to get it to work in native windows, with enough tweaking
Thank you kindly, will try this.
i love you guys
Thanks. There is hope.
It's awesome. Sets the stage for digging into possible techniques of training and fine tuning existing models (after they are uncensored). Would like more of this type and maybe an update on new things learned about it. I would like to be able to create custom LLM for various applications and use them in LLM applications. Need all the background I can find.
I do completely agree with you.
Big authors like OpenAI, they are forced to censor their language models, but personally I use the local uncensored model.
This article is literally the textbook definition of "for educational purposes" . I find you to be very unique and dynamic , which I like a lot ! The article is very knowledgeable indeed !
Good job Eric! It is very important for society to develop an understanding of what is happening, what is being done and why.
Amazing information
Excellent, please keep on keeping on. A 'newbie-dummy' style page would help us uneducated ones get to grips with this exciting adventure.
Thanks for the great article. We were giving this a try and encountered an error related to “utils.py”, in statement “from openai import openai_object”. This object is not there in recent “openai” package (we tried 1.14.3 which was installed as per the process). Do let us know on how do we could resolve this. Thanks
I enjoyed reading this, it really is a great article. I love how you included a play by play of each step that was needed in order to achieve the final model. If you ever have time, the dolphin-mixtral model is another of yours I would like to follow the process of creating
Great stuff , i just start dive into it.....great man