Train own custom dataset on GPT-4

#1

Post by spartanliving4u » Mon, May 15, 2023, 3:19 AM UTC

Recently an influencer claimed to have made a "companion chatbot"

https://www.msn.com/en-us/money/other/snapchat-influencer-launches-an-ai-powered-virtual-girlfriend-to-help-cure-loneliness/ar-AA1b7csm

can you already train your own custom dataset on GPT-4?

How was this done?

Re: Train own custom dataset on GPT-4

#2

Post by neom » Mon, May 15, 2023, 3:57 AM UTC

There are a bunch of them now

https://github.com/jerryjliu/llama_index is one

Re: Train own custom dataset on GPT-4

#3

Post by spartanliving4u » Mon, May 15, 2023, 4:38 AM UTC

neom wrote: post #2

There are a bunch of them now https://github.com/jerryjliu/llama_index is one

but this one also uses openAI

Re: Train own custom dataset on GPT-4

#4

Post by neom » Mon, May 15, 2023, 5:05 AM UTC

spartanliving4u wrote: post #3

neom wrote: post #2

There are a bunch of them now https://github.com/jerryjliu/llama_index is one

but this one also uses openAI

Probably doing this: https://platform.openai.com/docs/guides/fine-tuning

Re: Train own custom dataset on GPT-4

#5

Post by francylisboa » Mon, May 15, 2023, 10:25 AM UTC

I particularly dislike using the phrase "trained on your own data". We're not actually training anything. We're just showing a pre-trained and complex auto-completer how we'd like the outputs to talk about. Training involves a recalibrate the weights of those beast LLMs.

Re: Train own custom dataset on GPT-4

#6

Post by Anil1331 » Mon, May 15, 2023, 5:02 PM UTC

But GPT-4 has a context limit too even though it's big enough