How to Fine-Tune GPT 3.5-Turbo

Learn the four steps to fine-tune GPT 3.5 Turbo using Python and LangChain:
1. Prepare your data
2. Upload your data
3. Train your model
4. Use your model
For copy/pastable code check out:
haihai.ai/finetune

[@Branstrom]

Since I had already seen two similar videos I fast forwarded to the end just to see what you had trained it to do. I was like hm the meaning of life... Totally thought it was going to answer with 42. Apparently we have different holy books!

[@jordan-jones]

This is awesome. Detailed but to the point. Thanks for sharing, looking forward to more AI videos

[@hosseinbadrnezhad2019]

You answered many of my questions in this video

[@joao.morossini]

Thanks for the great video. It will be very useful

[@gcash3074]

Love it. Thanks for this video!

[@power-of-ai]

thank you for sharing your knowledge

[@Teawos98]

bro thank you so much for this video

[@haihailabs]

My pleasure

[@Vincent-mx4rk]

thanks for sharing, great job

[@barryhoffman9956]

Awesome! You feed it the book of Proverbs and it spits out Ecclesiastes.

[@tobiasabdon]

Great work!

[@haihailabs]

Thank you!

[@pypypy4228]

Thanks! Is number of epochs something you want to choose? If so, how many should I choose and how to come up with this number?

[@huilinghuang7811]

nice to have this feature

[@haihailabs]

Agreed!

[@LeonardoBoz]

nice man! Good video.

[@LeonardoBoz]

I am creating a assistant to my whatsapp. Do you think that i can train a fine tune model with all the products that i have on my database?

[@NamasteMax]

What I'm really trying to figure out, is there a way to successfully train it on newer documentation, like say newer library docs like React? So when I'm using it to help me code I get more up to date syntax?

[@p0gue23]

Do you need the step of generating a synthetic user question for each message, or can you leave the user strings blank, if, like in the present example you are training on a bunch of text and not an actual chat?

[@Teawos98]

legend

[@pypypy4228]

Liked and subscribed

[@rajutukadya2919]

Hi. Very informative video. I had a quiestion. You have used langchain here, but when we don't use langchain and directly use openai, what should the system prompt be when using the fine tuned model?

[@haihailabs]

System message is optional.

[@kingturtle6742]

Can the content for training be collected from ChatGPT-4? For example, after chatting with ChatGPT-4, can the desired content be filtered and integrated into ChatGPT-3.5 for fine-tuning? Is this approach feasible and effective? Are there any considerations to keep in mind?

[@astera-pt9je]

Do you need distinct user questions for each line of the jsonl file? I am preparing my data and most of the lines in the file contain "can you provide more details on this?" for the user content. In addition, if I remove entries with duplicate user questions, the data shrinks significantly which I think might be an issue since we should have as many lines as possible in our dataset.

[@haihailabs]

You’ll get better results with distinct questions

[@oskarrost9774]

Hi i have a quick question... I prepared my data but iam getting an error while launhing the upload file saying that in Line 1 of my data there is no dictionary, even tho in line 1 i just have the message system text. Do you know how to fix that?

[@XWSshow]

Do I Need an paid Account for Fine Tuning? Because I tried it out today, and I got an error, that fine tuning cant be created with an exploring Account…😢 (I still have 4 Dollars on my exploring Account)

[@Antonego64]

which Python version do you use in this video?

[@user-qi9np9jk3r]

Amazing video mate, you know i-v only one issue while doing this. While running FineTuningJob, the terminal returns this error: 'openai' has no attribute 'FineTuningJob' Which version of openai are you currently using?

[@haihailabs]

Make sure you pip install openai -upgrade to the latest version

[@trackerprince6773]

What is the difference between custom gpts vs fine tuning gpts?

[@v3teff]

I will receive an email regarding the account that I use to access the OpenAI key?

[@echofloripa]

I did a test using questions and answers about the elections process in Brazil. It had 67 questions and answers. I tried the default 3 epochs, 5, 7 and even 12. In none of the cases I managed to get the same response I had trained on, for exact same system message and user message. I tried in Portuguese and English language, and the result was the same. Yes, it gave a different response compared to the base model, but yet, never a correct answer. For the English dataset test I trimmed the 67 questions to only 10. You can check the loss of the training using its api and the numbers was erratic. I guess that at least in gpt3.5-turbo fine tuning, it's not possible to get it increase it's knowledge. I did some tests with open-source llms, but I still have to train with llama2. Maybe fine-tuning isn't really fit for that, and you have to use embeddings and vector databases to achieve that.

[@luiseduardo2249]

cara estou tendo problemas em adicionar o training data e o validation data do fine-tuning, ele pede em formato de prompt-completion, porem sempre que tento adicionar ele diz q esta em prompt-completion e precisa estar em chat-completion, oq acaba criando um loop de problemas, poderia me mandar algum modelo de validação e treinamento que saiba que funciona para eu testar?

[@echofloripa]

@@luiseduardo2249 mas deu erro nas chamadas da api da openai? Qual erro exatamente?

[@luiseduardo2249]

@@echofloripa estava com um erro da plataforma msm, resolvi. saberia me dizer se há como enviar de alguma forma um input para o gpt (configurado com fine tuning) do power automate por exemplo, usando API, apenas vejo a glr usando chat bot

[@echofloripa]

@@luiseduardo2249 faz tempo que não mexo na integração com a openai, comecei um projeto pessoal de App mobile educacional em Flutter que não deu tempo para mais nada. Não conheço este power automate, mas imagino que possa programar uma integração.

[@farahimad4432]

How do you get an email suggesting that your file is ready?

[@JackVucivic]

Thank you. Good to see you are using Proverbs. Very good Book of the Bible.

[@shubhamtyagi6281]

I have OpenAI key, but I'm not able to follow your video, can you make some detailed blog, share snippets, I would like to learn and create something similar in my cultural sacred book. I'm a coding beginner. I follow Udemy videos, etc. However I am very new to AI stuff. It would be very helpful if you did that Thanks.

[@haihailabs]

Sure thing! haihai.ai/finetune

[@timduck8506]

GPT 3.5 is floored as it is censored. so you need to find out were it is floored before you can ask the question which make's it 3 to 8 times slower to get the right answer you wanted if it comply's to the rules that gpt 3.5 go's by. else your going to get a error message or some thing that evades the answer to same face.

[@sahil91172]

Hey Brother , really nice video! I was wondering if I could help you edit your videos and also make a highly engaging Thumbnail which will help your video to reach to a wider audience .

[@Shahid_An-AI-Engineer]

Sir would you please share this fine tuning dataset with me as I'm a student and i don't have a access to GPT-4. If you wanna help please reply.

[@TheGuillotineKing]

Fun Fact you can fine tune but you can't get the model so they keep you paying for the API your better off training locally on your own machine

[@D3ADPIX]

Good luck deploying something as powerful at scale at a cheaper price. If you have any luck let me know.

[@TheGuillotineKing]

@@D3ADPIX hugging Face will let you fine tune for free on a small dataset and for 20 a month you get a lot of resources you should look into it

[@johan-mattias]

nice clock brah 🕰