Skip to content

[feature] Support SetFit: few-shot fine-tuning of Sentence Transformers, works with about 8 samples per class. #1633

Open
@stargazer33

Description

@stargazer33

It would be great to add SetFit to postgresml

See
https://pypi.org/project/setfit/
https://huggingface.co/docs/setfit

SetFit for text classification is different from other libraries: Usually, to train/fine-tune a model you need thousands of samples per class. In this example
https://postgresml.org/docs/open-source/pgml/guides/llms/fine-tuning
the "train" part of IMDB dataset contains 25K rows. There are 2 classes, so 12500 samples per class.

Now I'm quoting the SetFit documentation

It achieves high accuracy with little labeled data - for instance, with only 8 labeled examples per class on the Customer Reviews sentiment dataset

The code where they train a classifier - again they classifying film reviews (nothing really new here)
is here
https://huggingface.co/docs/setfit/main/quickstart#training

the sample_dataset function will sample only 8 samples for each class.

Compare this:
12500 samples per class
vs
8 samples per class with SetFit

In the real life, in many cases, you can collect... 50 samples per class and use SetFit to train a model.
Situations where you have tens of thousands of samples are quite rare.
Let's support SetFit.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions