Predictors and Models

Predictors

The base class which enables predictions in superduper is the Predictor mixin class.

A Predictor is a class which implements the .predict method; this mimics .predict from Scikit-Learn and related frameworks, but has support for prediction directly via the Datalayer.

A typical call to .predict looks like this:

predictor.predict(
    X='<key-column>'     # key of documents or column of table to take as input
    db=db                # `Datalayer` instance, built via `db = superduper()`
    select=my_select     # database query over which to compute outputs
    **predict_kwargs     # additional parameters for `.predict`
)

Examples of Predictor classes are the AI-API classes in

• superduper.ext.openai.OpenAI*
• superduper.ext.anthropic.Anthropic*
• superduper.ext.cohere.Cohere*

Models

A model is a particular type of Predictor which carries large chunks of data around in order to implement predictions. These blobs can be, for example, the weights of a deep learning architecture or similar important data.

Examples of Model are:

• superduper.ext.torch.TorchModel
• superduper.ext.sklearn.Estimator
• superdueprdb.ext.transformers.Pipeline

Each of these inheriting classes also implements the .fit method, which re-parametrizes the class in question, typicall via a machine learning task and objective function.

A typical call to .fit looks like this:

model.fit(
    X='<input-key-column>',    # key of documents or column of table to take as input
    y='<target-key>',          # key of documents or column of table to take as target of fitting
    db=db,                     # `Datalayer` instance, built via `db = superduper()`
    select=my_select,          # database query for training and validation data
    **fit_kwargs,              # additional parameters for .fit
)