Is it possible to add training data to an existing fine-tuned model?

[coral jasper]

I'm trying to fine tune a model on my dataset, but I think I would want to add more data to the trained model as I go along or as my product grows. How could I do that? Or should I just create new model with all the additional data every time I want to bring a change?

[hallow orchid]

Dependent on how much data you want to add, you need to decide between creating a system message if it's only one or two things you need to notify the model of, but if it's several fine tuneds you can create a new fine tuned model based off your current fine tuned model.

In other words, you can "fine tune a fine tuned model"

This way you aren't paying for a brand new model from scratch

[coral jasper]

So when I instruct it to fine-tune using the API. In model value, I can just pass in my model, instead of their basic models. Right?

[hallow orchid]

https://platform.openai.com/docs/guides/fine-tuning/continue-fine-tuning-from-a-fine-tuned-model

[final vessel]

I think the larger question is why would you fine-tune

[hallow orchid]

For their own use cases? Lots of people fine-tune.

[final vessel]

Which I think is a huge mistake

[hallow orchid]

Great, so you don't have to fine-tune, but lots of people have special use cases that would require it.

I guess that only applies to people with actual projects though

[final vessel]

I guess legal AI isn't an actual project

[hallow orchid]

I guess you think your use case is the only use case?

There are lots of different use cases. That's why fine-tuning exists - it helps tailor it to specific use cases. You are free to look at the documentation on it and why that might be necessary.

To claim it's a "huge mistake" is ridiculous and, quite frankly, ignorant of how this works.

[final vessel]

I do full-time development, a lot of it's consulting for other companies and still haven't ran into a use-case where it was needed

[hallow orchid]

Wow, good for you man! That's totally rad!

I kind of don't believe you, since anyone who is capable of developing understands that there's no way they could run into every, or even most, use cases, but cool!

So in the future, please don't muddy up someone's support request so you can give your opinion on something irrelevant!

[final vessel]

I don't think you know how fine-tuning works. Go try it and report back 👍

[hallow orchid]

I...do actually. I have my own fine-tuned model.

It handles specific case information that it needed fine-tuning on to ensure appropriate and accurate responses.

Otherwise its responses were quite sporadic in accuracy.

The point is, your opinion is just that - an opinion. And no one asked for it. So please don't come into people's support tickets and tell them that they're making a huge mistake based on your own experience with unrelated projects.

[final vessel]

He said he wants to add more data, I'm just trying to save him from the headache

He is just trying to teach a model new information which is a lost cause

Because that's not the purpose of fine-tuning

[hallow orchid]

First off, he never said he was trying to teach the model brand new information. Second, the entire purpose is few-shot learning, which helps provide more accurate examples without needing to include them in a prompt. You're assuming he doesn't know what he is doing. That's not appropriate, and is quite rude.

Also, if you read, he already has a fine-tuned model, he just needs to fine-tune it again. That's what he was asking about.

He did not ask our opinion on if he should.

[final vessel]

I literally just asked what is the reason you would fine-tune? I'm just trying to make sure he understands it before he spends thousands of dollars making his own custom model

[hallow orchid]

You came in saying it was a "huge mistake" without even knowing a use case, because in your own super professional and perfect mastery experience, you've never had to do it, so that must mean it's never required.

[final vessel]

cool.

[hallow orchid]

Also, not sure where you're getting thousands of dollars from.

Even if you had 1k tokens per example, you'd need to have over 33,000 examples for it to cost even $1000

er

I did my math wrong

Oh, no I didn't. Yeah, about 33,000 examples. Which, if that's your data set, I'm thinking you don't know how to fine-tune.

Please do not spread misinformation.

[brisk yarrow]

Hey there, @final vessel and @hallow orchid Hope you both are doing well.

I noticed that there's been a lively discussion between the two of you regarding fine-tuning models. It's great to see such passion, but I'd like to suggest taking the conversation to the DMs if you feel it's worth continuing.

While it's valuable to share different perspectives, let's remember to keep conversations respectful and professional. We're all here to learn and support each other. If you find common ground or come to a resolution, feel free to share your insights with the larger community later. Thanks! :)

[final vessel]

Hey thanks! And yeah no worries, I realised the same and decided to use the block function, all is good now! Just trying to help the original poster in DMs now instead Hope you have a good day Smile

[coral jasper]

@hallow orchid Thanks so much for holding on to your argument. I learnt quite a few intricate details just by reading your messages.

[coral jasper]

Thank you so much ✌️