Required Inputs

At minimum, Sibyl-API requires the following inputs, which can be provided as csv files or pandas DataFrames.

entities: A table with the entities to be explained. Each row should correspond to a single observation.

Columns:

• eid (required): unique identifier specifying which entity this observation corresponds to

• row_id: unique identifier specifying the observation ID. Together, eid and row_id should uniquely identify each observation.

• label: the ground-truth label for this observation

• [FEATURES]: additional columns for each feature used to make predictions. These columns should be named the same as the features used in the model.

Sample table:

eid	row_id	label	size	has_ac	nghrbh
house1	2001	10020	100	False	Oceanview
house2	2021	20304	120	True	Oceanview
house2	2001	56000	304	False	Oakvale

features: A table with the features used to make predictions. Each row should correspond to a single feature.

Columns:

• name (required): the name of the feature

• type (required): the type of the feature. This can be categorical, numerical, or boolean

• description: a description of the feature

• negative_description: a description of the feature when it is not present. Only for boolean features

• values: a list of possible values for the feature. Only for categorical features.

Sample table:

name	type	description	negative_ description	values
size	numerical	size in square feet
has_ac	boolean	has air conditioning	does not have air conditioning
nghbrh	categorical	neighborhood		[Oceanview, Ridge, Oakvale]

realapp: A Pyreal RealApp object fit to the ML task. See the Pyreal guide for more information on setting this up. For simple applications, you can generate a RealApp object from your model/pipeline and data with:

from pyreal import RealApp

# your_model: trained estimator or pipeline; any object with a .predict()
# your_training_data: pandas dataframe, training data for explanations

realapp = RealApp(your_model, x_train_orig=your_training_data)