Inefficient Fine Tune training

[warm fjord]

I have created a fine tune with the davinci base model and we have loaded it with legal information from Peru in PDF format: https://drive.google.com/file/d/1DEyjc_gu8tfTahNtD_RaysNL_34ZcL5W/view?usp=sharing

However, when validating the finetune, the answers it gives are not exact with the information that has been trained.

We need papers that contain the best practices to train your tuning or recommendations to improve our training and that the answers are more precise. 🙌

Fine Tune ID: davinci:ft-biwit:64136a1c1bdc9c77eaa4f784-1678995398564-2023-03-16-20-06-45

[atomic wedge]

You don't use fine-tuning for teaching it new knowledge

[warm fjord]

@atomic wedge So what do you recommend to use to have my own model and train it with information that the base models have not learned (the information comes from pdfs files)

[atomic wedge]

You have to use embeddings

[warm fjord]

@atomic wedge Do you have any guide on how to do it? thanks for answering

[atomic wedge]

Here is a basic guide: https://github.com/openai/openai-cookbook/blob/main/examples/Question_answering_using_embeddings.ipynb

[warm fjord]

Thanks @atomic wedge !

@atomic wedge My information is not limited to 1 document, there are thousands of documents, I do not see it as viable to follow that method 😥

[atomic wedge]

You can do it with thousands

[warm fjord]

@atomic wedge Do you have any reference of doing it with Node.js?

[atomic wedge]

I mostly do backend and use Python for that, what kind of structure are you working with?

[distant vigil]

How to save those embeddings to use in React or Node? Hdf5 with jsfive or Pickle? I have already created embeddings with h5 and able to work in python+ streamlit.

[warm fjord]

@atomic wedge Graphql, Apollo server , node.js, next.js, typescript