Fine tune update model not work properly

API
, , ,
1

When you fine-tune an existing model, do we send a new JSONL with just the new prompt-completions or we add the new prompt-completions to the same previous file (which has all the previous + the new ones)?

I tried with just the new ones, and it seems it completely forgot about the old ones. I also tried adding the old ones to the new one, and the cost it showed me was for the whole file (old + new). What am I missing :question:

2

How are you specifying the existing model when you come to retrain? can you include the code you used for this please.

3

@Foxalabs below is process and reference of API that we use

we follow this doc of open AI that mentions to update the fine tune model need to create a new model with a new data set and add the base model as the existing fine-tune model name:

below is API

curl https://api.openai.com/v1/fine_tuning/jobs \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{
    "training_file": "new training data",
    "model": "existing model of fine tuning"
  }'
4

Ok, and in the “existing model” name are you putting “gpt-3.5-turbo” or something different?

5

@Foxabilo

No the previous fine-tuned model that I created was based on the gpt-3.5-turbo model.

6

@Foxalabs Let me explain the whole process

  1. I created the fine tune model with 50 examples based on the
curl https://api.openai.com/v1/fine_tuning/jobs \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{
    "training_file": "file-abc123",
    "model": "gpt-3.5-turbo"
  }'

This gives me fine tune model and it works perfectly.

  1. Now I want to update the training data of this model so for that open ai doc and tell me that create a model based on this model so create the new 50 training data.
curl https://api.openai.com/v1/fine_tuning/jobs \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{
    "training_file": "new training data",
    "model": "existing model of fine-tuning that create in first steps"
  }'

in this case, old training data does not work properly. in this updated model.

1 Like
7

That seems to follow the correct format, if you are not getting the results you expect then I can only suspect that 50 is an insufficient number of examples to train with and the new training data is having more influence than the old or you may need to adjust the first training runs number of epochs, trying increasing them by 2, but typically additional fine tunes are only small amounts on top of a large existing set. If both sets are of a similar size then the results may be less than ideal.