API Fine-tuning Word Character Limit | OpenAI | Page 1

stoic valley Mar 2, 2023, 5:43 PM

#

Hi everyone; so my team is working to train the API model to read long government contracts and spit back out necessary/significant information so we would be able to read such contracts efficiently. When using the data-preparation tool on API, I found that the model wants us to limit our entries to <2000 tokens -- has anyone else conducted projects in which they had to work around a character limit? Also, if so, are there any common errors/edge-cases that you've encountered with the data-preparation tool that I could run into when training this data? Thank you

night mesa Mar 7, 2023, 3:24 AM

#

stoic valley Hi everyone; so my team is working to train the API model to read long governmen...

This is a fairly regular problem when trying to use GPT3/3.5 for analysis/summarization of large documents like contracts. You cannot work around the token limit but can work on a chunking the document appropriately and recursively extracting/abstracting each chunk. This is an evolving area and we'd love to understand as well.

smoky dawn Mar 7, 2023, 2:55 PM

#

How can I API Fine-tuning with chatGPT-turbo0301?
There is any JS code that do that?

echo bough Mar 7, 2023, 10:01 PM

#

smoky dawn How can I API Fine-tuning with chatGPT-turbo0301? There is any JS code that do ...

That model is not available for fine tuning

smoky dawn Mar 8, 2023, 5:53 AM

#

So which one?

echo bough Mar 8, 2023, 2:55 PM

#

https://platform.openai.com/docs/guides/fine-tuning/what-models-can-be-fine-tuned

wintry idol Mar 9, 2023, 10:47 AM

#

I'm also encountering the same issue. I tried chunking my text into 1500 words per chunk, then send it to ChatGPT.

I then store the response and then adjust the prompt into something like,

system: "Summarize someething. This was your previous answer. {previousAnswer}"
user: "Continue summarizing for the next text: {nextText}"

But the result is quite bad.

#API Fine-tuning Word Character Limit