Do we really need to fine tuning GPT3?

PinguSpy · Sep 13, 2022

The reason why we fine-tuning is to narrow down gpt3 to focus on certain pattern, format, or style of writing, thus.. reducing the cost.
Fine tuning also means discarding other possibilities.

Sometimes I think it doesn't make sense when you fine tune your model with thousands of articles.
Because gpt3 is already smart with hundreds of GB of data.
You don't need to feed it with another articles. Enough with good prompt engineering setup.

I think good prompt engineering setup is the key.

Is it?

splishsplash · Sep 14, 2022

PinguSpy said:
The reason why we fine-tuning is to narrow down gpt3 to focus on certain pattern, format, or style of writing, thus.. reducing the cost.
Fine tuning also means discarding other possibilities.

Sometimes I think it doesn't make sense when you fine tune your model with thousands of articles.
Because gpt3 is already smart with hundreds of GB of data.
You don't need to feed it with another articles. Enough with good prompt engineering setup.

I think good prompt engineering setup is the key.

Is it?

You dont fine tune it to write articles. It can already do that. You fine tune it to do more specific tasks like answering questions, summarization or whatever other specific task you want to try.
To teach it new knowledge to write about you dont fine tune. You just feed it the right data and potentially fine tune it for the type of summarization you want

Mactronn · Sep 14, 2022

My AI sites are doing well without fine-tuning. I've lost some traffic due to Google content update, but it's not alarming situation.

PinguSpy · Sep 16, 2022

So its more like pattern and style e.g template

ai_is_fun · Sep 18, 2022

PinguSpy said:
So its more like pattern and style e.g template

Yes, getting it to write full 1k-2k words in consistent style is hard. I've had major success finetuning for heading sections with dataset in the hundreds of thousands, you may want to look into doing it that way.

Pam Danny · Sep 29, 2022

ai_is_fun said:
Yes, getting it to write full 1k-2k words in consistent style is hard. I've had major success finetuning for heading sections with dataset in the hundreds of thousands, you may want to look into doing it that way.

So how is this done? Is it to find the articles already written in the format (template) that you like and sending it to the AI? And doing that in large quantities?

How is it done for someone that doesn't know how to code?

Akaroth · Jan 6, 2023

splishsplash said:
You dont fine tune it to write articles. It can already do that. You fine tune it to do more specific tasks like answering questions, summarization or whatever other specific task you want to try.
To teach it new knowledge to write about you dont fine tune. You just feed it the right data and potentially fine tune it for the type of summarization you want

Do I need to know Python and TensorFlow to do that?

splishsplash · Jan 6, 2023

DaniloRobert said:
Do I need to know Python and TensorFlow to do that?

Python, yes. TensorFlow, no.

HyperLesso · Jan 6, 2023

Unless you need a more specific article, you will need to do it. But for higher quality, it's no need(personal opinion)

paisawaisa · Jan 11, 2023

What is the process to fine tune it? Do we have to learn coding for this?

howard981 · Jan 11, 2023

paisawaisa said:
What is the process to fine tune it? Do we have to learn coding for this?

Yes. It's a little tricky for newbs. It involves scripting. Search for the offical OpenAI documentation for fine tuning and see if you can understand it or follow the steps.

Mactronn · Jan 11, 2023

howard981 said:
Yes. It's a little tricky for newbs. It involves scripting. Search for the offical OpenAI documentation for fine tuning and see if you can understand it or follow the steps.

It's not tricky at all! Anyone can copy the code from their website, edit the API key, set the model, import the dataset and run the code from terminal.

howard981 · Jan 11, 2023

Mactronn said:
It's not tricky at all! Anyone can copy the code from their website, edit the API key, set the model, import the dataset and run the code from terminal.

Not really. It would be hard for a newbie that doesn't even know what a terminal is. Many people would look at that doc and watch it cluelessly and be intimidated. Not all people are at the same computer skill level. A lot of people don't even know to navigate the docs.

paisawaisa · Jan 11, 2023

howard981 said:
Not really. It would be hard for a newbie that doesn't even know what a terminal is. Many people would look at that doc and watch it cluelessly and be intimidated. Not all people are at the same computer skill level. A lot of people don't even know to navigate the docs.

Reading from the instructions, I might be able to do it. Will try it. Any problem, I will ask chatgpt.

howard981 · Jan 11, 2023

paisawaisa said:
Reading from the instructions, I might be able to do it. Will try it. Any problem, I will ask chatgpt.

Good plan. I also found that ChatGPT doesn't give full instructions (as in step by step) for technical requests. So consider Googling technical blogs if you get stuck, or ask here. But do try to accomplish it on your own to gain more experience.

Arkus · Jan 11, 2023

howard981 said:
Not really. It would be hard for a newbie that doesn't even know what a terminal is. Many people would look at that doc and watch it cluelessly and be intimidated. Not all people are at the same computer skill level. A lot of people don't even know to navigate the docs.

It's very easy actually. I was playing with fine-tuning GTP-3 model without any coding skills. Their official documentation is very good so anyone can start using their API in a matter of time. Nothing tricky, nothing hard...

howard981 · Jan 11, 2023

Arkus said:
It's very easy actually. I was playing with fine-tuning GTP-3 model without any coding skills. Their official documentation is very good so anyone can start using their API in a matter of time. Nothing tricky, nothing hard...

I get it. But you observe it from lenses of an experienced user.

Look at this example from their docs:

We recommend using our OpenAI command-line interface (CLI). To install this, run

pip install --upgrade openai

(The following instructions work for version 0.9.4 and up. Additionally, the OpenAI CLI requires python 3.)

Not many people would understand where and how to run such command.

And some wouldn't even understand what "requires python 3" means.

If they were really spoon-feeding, they'd have links or instructions on how to run a terminal emulator for Windows and Mac and how to install python.

Like I said, people are at different computer skill levels and IQs.

I am a little new here, so maybe it's that users of this forum are closer to a power-user.

Adkins · Jan 11, 2023

It's more about fine tuning your prompts instead of fine tuning the model ;-)

PinguSpy · Jan 11, 2023

paisawaisa said:
What is the process to fine tune it? Do we have to learn coding for this?

https://www.blackhatworld.com/seo/s...o-code-real-method-practice-examples.1424188/ (learn how to)

https://www.blackhatworld.com/seo/s...ht-now-fine-tuned-model-any-use-case.1440319/ (prompt)

But i dont have time for that, I end up hiring an expert to help me.

It took him 6 months, to come out with 1 click 1 article (850 words+) without repetitive paragraphs issue.

This task almost feels like "article structure engineering". Lmao but It's a fun thing to do.

Do we really need to fine tuning GPT3?

Elite Member

Elite Member

Power Member

Elite Member

Newbie

Elite Member

Senior Member

Elite Member

Regular Member

Elite Member

Newbie

Power Member

Newbie

Elite Member

Newbie

Newbie

Newbie

Regular Member

Elite Member

Main Menu

Marketplace

Making Money

BlackHat World