CJ Gammon
A quick walkthrough of training a fine-tuned model on gpt-3 using the openai cli.
In this video I train a fine-tuned gpt-3 model on Radiohead lyrics so that it will generate a Radiohead song form the first line.
Documentation: https://beta.openai.com/docs/guides/fine-tuning
I am sorry, but I am having a huge problem that I can't seem to find anywhere. It's just crazy to me that all of these tutorials feel the need to explain API KEYS, don't we all know what those are? On the other hand none of the tutorials seems to explain how you are issuing the commands. I am trying to use the command prompt window. I have everything installed. I have got all the way up to the point of trying to use the prepare data tool, and I can't get any further. Its just crazy to me that everyone's videos feel the need to explain API keys, like the video is for amateurs, but in reality its not. I'm an amateur, I know API keys. What I don't know is how to run these command windows.
Awesome video ! Thanks a lot. Question – If I build a Health related application and fine tuned with health related data, would users still be able to generate essays, poetry and non health related content from the base model ? Or would this restrict the responses strictly to health ? Thanks for your help !
And this is extremely expensive
So you pay $8 every time you fix any issue and have to recreate the model? That is ridiculous
Thanks for this video ! How did you do to prepare that much DATA ? I have several huge files of DATA to prepare, and I'd like to know how to prepare them correctly ?
Thanks !
The way you're easily just going over the docs is just amazing. It's as if you wrote the documentation yourself. No hurries, no shortcuts……Just purely methodical
did anyone face this issue and how did you solve it ?
Stream interrupted (client disconnected).
One doubt.Do we need to pay only during training? And accessing that gpt later not cost incurred?
Can you plz tell how can i use the same to bild my apis ?? and fetch the repsonse via apis
Just FYI this had your email in it in the top right. Might not want that visible.
hey CJ, wondering how big is your data set? Looks like you put in quite a few examples. My understand is for LLM you don't need to put in that many examples for fine tuning. Have you tried putting in 5-10 examples? I wonder what the output is. Great video!
What graphics card and virtual memory are you using?
Thanks for making this video.
If you have an answer, I'm trying to figure out a question while doing research. The fine tuning seems to work well for conforming to patterns of speech for phrases, but would this work for limiting the output of a model to a set of words, while not directly influencing the patterns of speech used?
I'm planning to limit the output to utilizing only words from a set list of words. Ideally this would cause the model to simply find output patterns that work together while constrained only by choice of vocab.
Do you think if I structure the data in the JSON correctly, it will take a list of words as fine tuning data?
Hi cj, great job will you upload some more video where everyone can learn from you.
Admittedly, I haven't played much with this but what is the usefulness of this? There must be more to it. Seems like a spreadsheet would do just as well, but why did you need to exactly use the same arrow -> ending? What happens if you leave it off? I'm just asking if you can do anything useful with it, like, ask for the rest of the lyrics translated into French, just as an example. Sorry for the naïve questions, I'm not trying to be critical.
Great video, CJ! Question about the training data. Do I understand correctly that the dataset you prepped is sent to OpenAI? If so, do you know how do they store and if they use your data?
I have a question, when I prepared my data I used the right formatting as stated on the fine tuning instructions for example:
{"prompt": "What is your number", "completion": "I don't have any number"}
When I format the file to a .jsonl with the cli and open the file the instruction looks like:
{"prompt":"","completion":" What is your number", "completion": "I don't have any number"}"}
Is that the right format to train the model now ? Or did something get messed up in the formatting process to the .jsonl ?
Thanks
This is great! Any way we could see the data you sent? I'd be eager to try to reproduce your example!
is this ideal for websites such as FAQs, TOS, PP and product details? where a user will ask a question.
If my data is confidential can I do training locally without sending to openai?
hey, any tips on how to fine tune a model based on a very large pdf document without the "n" to split prompt/resolution? I thought maybe have a script break down in every question mark? Or is there some other way?
Thank you such a clear tutorial. I have a question though, how can I save the fine tuned model, so I can use it later in my application?
Cool, but I dont fully understand the benefit if you need to give the model answers to your prompts in advance? Could you not just query your own JSON file that you created? You send a prompt and if it matches you get back the answer inisde the JSON? Do you know what I mean? What did you really pay 8 bucks for here?
in the CLI it says "Make sure to include `stop=[". END"]` so that the generated texts ends at the expected place."
can someone please give me guidance about where exactly to place `stop=[". END"]`?
Hi sir thanks for this video! A question: how would you rate the performance of fune-tune model VS songs generated from other approaches like few-shots/zero-shots?
Hey CJ, can you share your python code being used for this model?
Great video, thanks!!!
before i try can i fine tune a novel that already have 210 chapters translated by humans can i use those to train the ai to translate the other 800 chapters?
That's exactly what I'm looking for, Thank you
Hey CJ , can you please share the dataset you have used and the utility script you have written on it.
THank you for the video, do you know how to fix the problem with "tream interrupted (client disconnected)."?
The part I am confused is that when fine tuning it completes my prompts perfectly, but when I try to have a conversation it doesnt remember the context of previous messages. So i got the idea to feed it back in as an input the entire conversation up to that point, however it gets all crazy after 3-4 messages because the training dataset does not contain any prompts that have conversation history in it, so how is this done in order to have conversations with it like chatgpt?? Please someone help me
first video of finetuning on a prediction task, many have been on classification.
What did you do when you got the prompt that your API Key was not working because i am stuck at this point and my export command doesn't work, i have tried using set instead of export and it still doesn't work
We are building a chatbot for our organization. We have thousands of different types of source documents using which the bot has to provide the answer to user queries. We could run the similarity search on the entire set of documents but that may not give the efficient results so we thought of making various clusters of different types of sources documents. For a user query, the similarity search should happen only to a cluster of documents where the query belongs to. The bot should be able to determine which cluster needs to be referenced to provide answer to user queries. Now, the question is how can we make the bot understand which cluster to refer for a particular user query.
We thought of creating a decision prompt template which will take the user query as input. We will feed some examples that can help LLM determine which cluster to refer. But the problem with this approach is that we won't be able to train the LLM efficiently using this approach for our data. The data source is very huge and we can't train the model for every type of question which a user may have from these data sources. Can someone please suggest that how should we approach this challenge.
Does anyone have any idea where I can find the dataset containing lawyer-based prompts?
I am using windows 10os..how to do this windows.
How to prepare this datset in json format on windows OS 10.
Link is broken. It redirects to main documentation with information tha fine-tuning is not available yet for Gpt3 and Gpt4.
I'm interested in the performance benefits of a large context window in training. Particularly be interesting if this window also gets larger if they allow to find tune 16k version. Does anyone have any experience with this?
Is it mandatory to use a structure prompt–> completion for training? Or can i just give some Jason with every input I want?
Do someone has an Solution for c# especially for unity?