Update the existing fine tuned model

API
1

Am trying to add some more prompts to the existing model, however when i do that the old one just goes away and I have to create a new fine tuned model which is not cost effective.

What i have tried till now

Original model:

openai api fine_tunes.create -t “test_prepared.jsonl” -m davinci --suffix “test”
Model created : davinci:ft-personal:test-2023-02-12-14-39-06

fine tune the existing tuned model

openai api fine_tunes.create -t “test_prepared_2.jsonl” -m davinci:ft-personal:test-2023-02-12-14-39-06 --suffix “test”

Model created : davinci:ft-personal:test-2023-02-<>
Newly created model works only with new data set and old data is ignored.

Am i doing anything wrong ?

2

The same, I have asked and complained about this last time. You have to pay the whole amount of data no matter how many examples you want to add to the previous training data. Fine-tuning an existing fine-tuned model does not mean “Update your existing data” , instead it aims to create a very new model that you have to pay from scartch. But it should not be like this.

1 Like
3

as per the document it says otherwise… but i doubt so… its not an amendment to the existing data set. At least thats what i discovered

4

It’s hard to tell unless you post some lines from your JSONL training file you used on your davinci:ft-personal:test-2023-02-12-14-39-06 model.

Please posts some lines from your JSONL file, as follows, wrapping your data in triple back ticks (markdown for code blocks).

```
# json line 1
# json line 2
# json line 3
# json line 4
# json line 5
```

Also, please provide all your fine-tuning params for that file as well.

Thanks!

:slight_smile:

5

Yes, it’s not an amendment, it creates another one. So if you want to include the newest examples, you have to fine-tune the whole data including the previous one again.

6

Thanks for the reply, basically the two command what i have list is how i created. Did not use any additional params.

And for the data, might not be able to post directly the data.
ex :
first jsonl → has use case specific data
second jsonl → this has general data prompts like Hi , hello and getting things.

Post fine tuning the existing tuned model… it just took greeting as base data and ignored and original data.

7

yeah this is what my experience as well… so all the time you end up spending more and more :slight_smile:

8

Yes, right! Even you want to add one line example, you have to pay for the whole data.

9

Hi @pravinvram

Sorry, but I cannot help you further if you do not post the data requested, exactly:

  • Your JSONL data, as many lines as you wish, but not more than 5
  • Your exact fine-tuning parameters you used when you created your fine-tuning.

Without this data, we are just wasting time, guessing.

:slight_smile: