Free Fine-Tuning extended to October 31st, 2024 by OpenAI

[PaulBellow]

OpenAI Developers (@OpenAIDevs) on X

Any fine-tuning projects you’re working on?

Feel free to share with the community!

[PaulBellow]

I’m going to fine-tune using my novels again, I think. I wonder if @eawestwrites has any recent tips?!

I might also use my 100,000+ character backstories to fine-tune.

What about you? Anything you’re working on fine-tuning?

Problems you’ve run into? Tips for us newbies?

[SpuriousGer]

I have all my emails from work that I would love to use as a base for a fine tuned model. They currently exist as .msg files. Anyone with an idea on how to approach this best?

[jr.2509]

As an input for the fine-tuning you need to create a data set composed of input-output pairs, i.e. pairs of user messages and assistant messages, as shown here in the example training data format, along with a system message (optional):

In your case you could construct this as follows:

• Input/user messages could be either a few bullets that outline the intended content of the email and the recipient OR an email from another person that you have responded to
• Output/assistant message would then be the corresponding email from your collection of emails that you have sent and that reflect the desired style of emails

In the system message you could provide some additional instructions regarding the task and some overarching stylistic guidelines. However, given you provide actual examples, you don’t have to overly detailed.

For additional guidance, you can take a look at the following fine-tuning guide.

[terekita]

Hello. Would you mind describing your process for fine-tuning models on novels? I had some success doing this using prompt/completion pairs with davinci, but now I’m trying to adapt for gpt-4o format. Thanks!

[jr.2509]

Hi @terekita - welcome back to the Forum!

Have you had a chance to look at this fine-tuning guide yet?
Let us know if you have a specific question and hopefully we’ll be able to point you in the right direction.

[terekita]

Thanks for your message. It’s more just a general question about what approaches people find useful for fine-tuning in the style of a novel. For my prior fine-tuning on davinci, I would take 3 sentences from the text and make that the prompt, and then the next three sentences for the completion.

For gpt-o mini I’m essentially doing the same thing for the user and assistant messages, and adding in a system message. This seems to work well also, but I’m wondering what other strategies people use for this task.
Thanks!

[jr.2509]

Here are a few links to posts you might find helpful:

Fine tuning the model for writing via famous author style

Right. I’m not OpenAI staff, but I’d still check. They might not want Stephen King’s books on their servers lest there’s a lawsuit. @Foxalabs laid out how to do it, though. Whether you do or not is up to you, but be careful. Better to be safe than sorry, as they say… Sometimes, though, simply telling it to write in a style inspired by Stephen King will get his basic style without having to fine-tune his content. Likely because there’s so much writing about Stephen King’s writing style on the i…

Ideal input / output form for Fine Tuning?

I am taking a several long books (biographies) and transforming them into input/output pairs to fine-tune GPT 3.5 on. My method is to transform the entire book into ~500 word “chunks”. Then tell the GPT 3.5 to look at each “chunk” and write a question that would plausibly elicit the chunk as a response. This will transform the book into input/output pairs. As far as fine-tuning, I’m wondering what the ideal “input” would be. A brief, generic question? A detailed, really specific question? A lo…

Prepare data for fine-tuning a story generator

Hey all. I have a dataset consisting of short children’s stories of a certain reading level. The stories look like this: Wake up May. It’s Monday It’s a school day. Come and eat breakfast. Good morning, Mom. What time is it? It’s seven o’clock. etc. I want to fine-tune a model to be the best it can be in generating stories for this level. How would I best prepare the dataset? Any thoughts? Thank you!