Version 0.1

Model Card Generator

This tool can be used to create a Model Card: a way to share information about a model’s intent, data, architecture, and performance. Use the form below to generate a markdown (*.md) file version of a model card.

Note: A red asterisk (*) indicates a required field.

Overview

What is your model name?*

Enter names of model owners in a "+" separated list.* Example: Jane Doe, Ph.D. + John Smith

Enter affiliated agencies in a + separated list.* Example: xD + Census

How was the model acquired?*

Anticipated Use

In which agencies will this model be used?*

Who are the intended users of the model?*

What are the intended use cases of the model?*

Model Information

What is the current model version?*

What is the version release date?*

What changes have been made since the last release?*

What is the license for use?*

Model Architecture

What type of algorithm is used?*

How is the input data formatted?*

How is the output data formatted?*

Datasets

Does the training dataset contain information related to individuals or human populations?*

What is the source(s) of the training data?*

How was this data collected or generated?*

What variables are contained in this dataset?*

How many entries are contained in your dataset?*

What percent of data is chosen as a validation set?*

Performance Metrics

What metrics are used to rate model performance?*

How are the metrics being reported?*

What is the confidence interval of these metrics?*

What decision threshold was used to compute the metric?*

What factors limit the model's performance?*

Bias

Does the training dataset contain information related to individuals or human populations?

Within the workflow of this model, is there risk of human judgement injecting bias?

What methods are used to minimize bias from human judgement?

What biases are potentially found in the training dataset from collection methods, sample size, representation, etc.?

What evaluation tools have been used to evaluate bias in the training dataset?

How and when are biases in the dataset addressed in the workflow of the model?

What testing has been performed to look for bias in the model?

What percentage of development time has been dedicated to bias mitigation?