Skip to content

Open Sentencing Model is a statistical model that estimates the size of criminal sentencing discrepancies due to race based on a range of demographic and legal factors.

License

Notifications You must be signed in to change notification settings

Call-for-Code-for-Racial-Justice/Open-Sentencing-Model

Repository files navigation

License Community Hacktoberfest

Call for Code for Racial Justice

A project for the 2020 Call for Code Challenge seeking technology solutions for racial injustice.

embraceLogo

IBM Cloud

IBM Cloud platform Apache 2

What's the problem?

Black defendants face harsher sentences than people of other races for similar offenses (e.g. harged at higher rates, assigned more significant charges, convicted at higher rates, given longer sentences, and denied parole more often).

How can technology help?

We built a statistical model that estimates the additional years of prison that a non-white defendant will face in a particular case. This tool allows public defenders and other criminal legal system actors to intervene in a particular case before an unequal sentence is handed down. The model also helps public defenders identify factors that lead most often to racial sentencing discrepancies.

Data Set

Cook County Sentencing data set used in the Open Sentencing model.

Roadmap and Architecture

This project will ultimately be presented as a component of a group of projects that will generate other predictions, capture new data, and generate reports. The architecture of the complete system is shown below.

Architecture Digaram

Machine Learning Algorithm

  1. Build a model to predict prison sentence length based on a range of factors including race and criminal charge.
  2. For a given charge, predict the sentence length. Then switch the race variable from non-White to White or White to Black and make the same prediction.
  3. Calculate the discrepancy between the predicted sentence length with the actual race and the predicted sentence length with the counterfactual race.
  4. Report the size of the estimated discrepancy in sentencing due to race and the relative severity of the discrepancy compared to past examples.

See the Notebooks folder in this repository for detailed Jupyter notebooks that demostrate how the model was trained and allow you to re-train the model yourself.

Create and deploy a Python Flask application

We have applications available for Node.js Express, Go Gin, Python Flask, Python Django, Java Spring, Java Liberty, Swift Kitura, Android, and iOS.

In this sample application, you will create a Python cloud application using Flask. This application contains an opinionated set of files for web serving:

  • public/index.html
  • public/404.html
  • public/500.html

This application also enables a starting place for a Python microservice using Flask. A microservice is an individual component of an application that follows the microservice architecture - an architectural style that structures an application as a collection of loosely coupled services, which implement business capabilities. The microservice exposes a RESTful API matching a Swagger definition.

Steps

You can build the application locally by cloning this repo first. After your app is live, you can access the /health endpoint to build out your cloud native application.

Building locally

To get started building this application locally, you can either run the application natively or use the IBM Cloud Developer Tools for containerization and easy deployment to IBM Cloud.

Native application development

Running Flask applications has been simplified with a manage.py file to avoid dealing with configuring environment variables to run your app. From your project root, you can download the project dependencies with (NOTE: If you don't have pipenv installed, execute: pip install pipenv):

pipenv install

To run your application locally:

python manage.py start

manage.py offers a variety of different run commands to match the proper situation:

  • start: starts a server in a production setting using gunicorn.
  • run: starts a native Flask development server. This includes backend reloading upon file saves and the Werkzeug stack-trace debugger for diagnosing runtime failures in-browser.
  • livereload: starts a development server via the livereload package. This includes backend reloading as well as dynamic frontend browser reloading. The Werkzeug stack-trace debugger will be disabled, so this is only recommended when working on frontend development.
  • debug: starts a native Flask development server, but with the native reloader/tracer disabled. This leaves the debug port exposed to be attached to an IDE (such as PyCharm's Attach to Local Process).

There are also a few utility commands:

  • build: compiles .py files within the project directory into .pyc files
  • test: runs all unit tests inside of the project's test directory

Your application is running at: http://localhost:3000/ in your browser.

  • Your Swagger UI is running on: /explorer
  • Your Swagger definition is running on: /swagger/api
  • Health endpoint: /health

There are two different options for debugging a Flask project:

  1. Run python manage.py runserver to start a native Flask development server. This comes with the Werkzeug stack-trace debugger, which will present runtime failure stack-traces in-browser with the ability to inspect objects at any point in the trace. For more information, see Werkzeug documentation.
  2. Run python manage.py debug to run a Flask development server with debug exposed, but the native debugger/reloader turned off. This grants access for an IDE to attach itself to the process (i.e. in PyCharm, use Run -> Attach to Local Process).

You can also verify the state of your locally running application using the Selenium UI test script included in the scripts directory.

Note for Windows users: gunicorn is not supported on Windows. You may start the server with python manage.py run on your local machine or build and start the Dockerfile.

IBM Cloud Developer Tools

Install IBM Cloud Developer Tools on your machine by running the following command:

curl -sL https://ibm.biz/idt-installer | bash

Create an application on IBM Cloud by running:

ibmcloud dev create

This will create and download a starter application with the necessary files needed for local development and deployment.

Your application will be compiled with Docker containers. To compile and run your app, run:

ibmcloud dev build
ibmcloud dev run

This will launch your application locally. When you are ready to deploy to IBM Cloud on Cloud Foundry or Kubernetes, run one of the commands:

ibmcloud dev deploy -t buildpack // to Cloud Foundry
ibmcloud dev deploy -t container // to K8s cluster

You can build and debug your app locally with:

ibmcloud dev build --debug
ibmcloud dev debug

Running the Examples and Calling the API?

Please find the notebook used to build the model [here](notebooks/Explore and Train - reader friendly.ipynb). Another notebook outlining how to make predictions with a pretrained model is here

If the service is cloud deployed be sure to point to the IP and port associated witht the deployment. Examples here are provided for deployment on localhost:3000.

You can then use Postman app to send json formatted POST HTTP requests to the model at the URL http://127.0.0.1:3000/predict Here is an example JSON data you can use to test the model.

{
	"PRIMARY_CHARGE_FLAG":false,
	"DISPOSITION_CHARGED_OFFENSE_TITLE":"[POSSESSION OF CONTROLLED SUBSTANCE WITH INTENT TO DELIVER\/ DELIVERY OF A CONTROLLED SUBSTANCE]",
	"CHARGE_COUNT":2,"DISPOSITION_CHARGED_CLASS":"2",
	"CHARGE_DISPOSITION":"Plea Of Guilty",
	"SENTENCE_JUDGE":"Maura  Slattery Boyle",
	"SENTENCE_PHASE":"Original Sentencing","SENTENCE_TYPE":"Prison",
	"COMMITMENT_TERM":3.0,"COMMITMENT_UNIT":"Year(s)", "LENGTH_OF_CASE_in_Days":336.0,
	"AGE_AT_INCIDENT":52.0,"RACE":"Black","GENDER":"Female","INCIDENT_CITY":"Chicago",
	"LAW_ENFORCEMENT_AGENCY":"CHICAGO PD",
	"LAW_ENFORCEMENT_UNIT":"District 25 - Grand Central",
	"UPDATED_OFFENSE_CATEGORY":"Narcotics"
}

From the Command Line

Run the following command line curl command

curl -X POST -H "Content-Type: application/json" -d '{"PRIMARY_CHARGE_FLAG":false,"DISPOSITION_CHARGED_OFFENSE_TITLE":"[POSSESSION OF CONTROLLED SUBSTANCE WITH INTENT TO DELIVER\/ DELIVERY OF A CONTROLLED SUBSTANCE]","CHARGE_COUNT":2,"DISPOSITION_CHARGED_CLASS":"2","CHARGE_DISPOSITION":"Plea Of Guilty","SENTENCE_JUDGE":"Maura  Slattery Boyle","SENTENCE_PHASE":"Original Sentencing","SENTENCE_TYPE":"Prison","COMMITMENT_TERM":3.0,"COMMITMENT_UNIT":"Year(s)","LENGTH_OF_CASE_in_Days":336.0,"AGE_AT_INCIDENT":52.0,"RACE":"Black","GENDER":"Female","INCIDENT_CITY":"Chicago","LAW_ENFORCEMENT_AGENCY":"CHICAGO PD","LAW_ENFORCEMENT_UNIT":"District 25 - Grand Central","UPDATED_OFFENSE_CATEGORY":"Narcotics"}' localhost:3000/predict

You should get the following result

{
  "model_name": "sentence_pipe_mae1.555_2020-10-10_02h46m24s",
  "sentencing_discrepancy": 0.211,
  "severity": 0.555
}

The model_name is the file name (with .pkl extension) of the model file used to make the prediction. Discrepancy and Severity are discussed in the notebook (Add notebook link here). See also Calling API and Comparing Results, for details on how the data is generated.

Next steps

License

This sample application is licensed under the Apache License, Version 2. Separate third-party code objects invoked within this code pattern are licensed by their respective providers pursuant to their own separate licenses. Contributions are subject to the Developer Certificate of Origin, Version 1.1 and the Apache License, Version 2.

Apache License FAQ

About

Open Sentencing Model is a statistical model that estimates the size of criminal sentencing discrepancies due to race based on a range of demographic and legal factors.

Topics

Resources

License

Stars

Watchers

Forks