Fine Tuning troubles, not working as expected

General API discussion
#1

I attempted to fine-tune some basic JSONL to test the waters.

{"prompt": "solar", "completion": "panel"}
{"prompt": "elon musk", "completion": "rules"}

I successfully trained an ada model. I say y to the suggestions. The fine tuned model was created successfully. However, when I tried testing it, I got the following results:

➜  Desktop openai api completions.create -m ada:ft-personal-2022-07-11-07-56-03 -p "solar ->"
solar ->

2-white

2-gray/white

2-%

and

➜  Desktop openai api completions.create -m ada:ft-personal-2022-07-11-07-56-03 -p "solar"     
solar sun, with the Sun and Earth being in a celestial system that is showing life%

and

 Desktop openai api completions.create -m ada:ft-personal-2022-07-11-07-56-03 -p "elon musk"
elon musk, the featured action hero of the two films and Ty main central character in the%

I thought whatever prompts and completions I would give should be given maximum weight. What is going on here? Why didn’t the model respond as it was fine-tuned to do? Also why do the results continually have typos (“and Ty main”)?

#2

At what temperature and spectrum?

#3

Oh… a good point. Maybe the default temperature isn’t maxed out. Let me try again.

1 Like
#4

How many samples did you use?

#5 
{"prompt": "solar", "completion": "panel"}
{"prompt": "elon musk", "completion": "rules"}

The two I provided.

#6

You need 200 samples minimum for finetuning.