#Using a Fine Tune Model

12 messages · Page 1 of 1 (latest)

worthy pebble
#

Hey Dimon, looks like you've made some good progress already!

#

Making a JSONL is right.

{"prompt": "How to open a sales report", "completion": "Step 1. Open the main menu; Step 2. Click on the sales report link"}

The data looks right but there is formatting missing.

# 

Each completion should start with a whitespace due to our tokenization, which tokenizes most words with a preceding whitespace.```
#

I also noticed that what you typed in How to open sales report
is not an exact match to the prompt you provided How to open a sales report

I'm not super experienced with creating fine tune models, and certainly not for when you want a very specific result out.

digital oar
#

@worthy pebble oh, thank you for the quick response! I will take a look at the format. Meanwhile, my goal is to train the machine on my documentation (we build an accounting software), which has basic FAQ and other instructions and articles. I'm far from understanding basics of gpt technologies and I'm not even sure that my goal is feasible. Eventually, I want to have a bot that will answer user's questions based on knowledge I provided. However, I see that current api give me some answers which can be irrelevant for my app.

fossil sequoia
digital oar
#

@fossil sequoia no, never heard about it. ChatGPT just impressed me so much that I was trying to use it right away )) But thank you!

fossil sequoia
#

You can still link OpenAI's API into your conversation, but for a pre-formed conversation with "topics", there's easier ways 🙂

Good luck!

digital oar
#

@fossil sequoia I'm sorry in advance for silly questions. If I use Fine-tuning, is it still possible that api will return an answer from "internet" which potentially can be wrong? Let's say user asks "How to generate a sales report" and api gives him some hint from another accounting app, not mine?

fossil sequoia