Their fine-tuning is only meant for formatting, not for injecting new data. However, I think you could use it to create an instruct or chat version of davinci
Fine-tuning works great, and like their documentation says, can result in much better output than zero shot prompt engineering. Their documentation says you need at least a few hundred examples. And also you need to choose a model that can actually handle the task you’re giving it. Finally, it is very important to choose the correct hyper parameters such as epoch and batch size during training. If you are getting gibberish back from your fine-tuned model, you did not do it correctly.
Honestly, this is comparing apples and oranges. The very powerful models of GPT-3.5 and GPT-4 are being compared to GPT-3. And then on that comparison, we are saying that custom fine tuning (which is currently not available on GPT-3.5 and GPT-4) is inferior or not useful?? Definitely use prompt engineering, but it's too early to call custom fine tuning garbage.
GPT-4 allows up to 32k tokens in a prompt. 1k tokens is about 750 words. So you could have 20,000 words of documentation in a single chatbot. If you wanted to try fine tuning though, someone else said you basically have to say the same things hundreds of times before GPT-3 understands that it needs to respond in that way. So maybe that would work. I'm just being honest about my personal experience. Fine Tuning hasn't worked for me. But prompt engineering had worked incredibly well if the prompts are written in the right way
@@NativeNotify but isn't spending some several thousand tokens each time a user starts using gpt-4 an expensive way? Is there a way to train your gpt bot by prompt engineering beforehand and all users use that pre-trained bot and you don't need to teach it every time?
![ปฏิเสธไม่ไหว (Crush on you) - LIPTA feat. No One Else [Official MV]](http://i.ytimg.com/vi/NBw1jF342vw/mqdefault.jpg)
Their fine-tuning is only meant for formatting, not for injecting new data. However, I think you could use it to create an instruct or chat version of davinci
This was very helpful information. Thanks!
Fine-tuning works great, and like their documentation says, can result in much better output than zero shot prompt engineering. Their documentation says you need at least a few hundred examples. And also you need to choose a model that can actually handle the task you’re giving it. Finally, it is very important to choose the correct hyper parameters such as epoch and batch size during training. If you are getting gibberish back from your fine-tuned model, you did not do it correctly.
Honestly, this is comparing apples and oranges. The very powerful models of GPT-3.5 and GPT-4 are being compared to GPT-3. And then on that comparison, we are saying that custom fine tuning (which is currently not available on GPT-3.5 and GPT-4) is inferior or not useful?? Definitely use prompt engineering, but it's too early to call custom fine tuning garbage.
But if you have a huge pdf documentation that you want your model to be trained, you cannot do this just with prompting.
GPT-4 allows up to 32k tokens in a prompt. 1k tokens is about 750 words. So you could have 20,000 words of documentation in a single chatbot.
If you wanted to try fine tuning though, someone else said you basically have to say the same things hundreds of times before GPT-3 understands that it needs to respond in that way. So maybe that would work.
I'm just being honest about my personal experience. Fine Tuning hasn't worked for me. But prompt engineering had worked incredibly well if the prompts are written in the right way
@@NativeNotify but isn't spending some several thousand tokens each time a user starts using gpt-4 an expensive way? Is there a way to train your gpt bot by prompt engineering beforehand and all users use that pre-trained bot and you don't need to teach it every time?