Innovation for Inclusion: Hack.The.Bias with Amazon SageMaker

This submit was co-authored with Daniele Chiappalupi, participant of the AWS pupil Hackathon workforce at ETH Zürich.

Everybody can simply get began with machine studying (ML) utilizing Amazon SageMaker JumpStart. On this submit, we present you ways a college Hackathon workforce used SageMaker JumpStart to rapidly construct an utility that helps customers determine and take away biases.

“Amazon SageMaker was instrumental in our undertaking. It made it straightforward to deploy and handle a pre-trained occasion of Flan, providing us a strong basis for our utility. Its auto scaling characteristic proved essential throughout high-traffic intervals, guaranteeing that our app remained responsive and customers obtained a gradual and quick bias evaluation. Additional, by permitting us to dump the heavy job of querying the Flan mannequin to a managed service, we had been in a position to preserve our utility light-weight and swift, enhancing consumer expertise throughout varied gadgets. SageMaker’s options empowered us to maximise our time on the hackathon, permitting us to deal with optimizing our prompts and app moderately than managing the mannequin’s efficiency and infrastructure.”

– Daniele Chiappalupi, participant of the AWS pupil Hackathon workforce at ETH Zürich. 

Answer overview

The theme of the Hackathon is to contribute to the UN sustainable objectives with AI expertise. As proven within the following determine, the applying constructed on the Hackathon contributes to 3 of the Sustainable Improvement Targets (high quality schooling, focusing on gender-based discrimination, and lowered inequalities) by serving to customers determine and take away biases from their textual content in an effort to promote honest and inclusive language.

As proven within the following screenshot, after you present the textual content, the applying generates a brand new model that’s free from racial, ethnical, and gender biases. Moreover, it highlights the precise elements of your enter textual content associated to every class of bias.

Within the structure proven within the following diagram, customers enter textual content within the React-based internet app, which triggers Amazon API Gateway, which in flip invokes an AWS Lambda perform relying on the bias within the consumer textual content. The Lambda perform calls the Flan mannequin endpoint in SageMaker JumpStart, which returns the unbiased textual content outcome through the identical route again to the front-end utility.

Software improvement course of

The method of growing this utility was iterative and centered on two predominant areas: consumer interface and ML mannequin integration.

We selected React for the front-end improvement on account of its flexibility, scalability, and highly effective instruments for creating interactive consumer interfaces. Given the character of our utility—processing consumer enter and presenting refined outcomes—React’s component-based structure proved excellent. With React, we might effectively construct a single-page utility that allowed customers to submit textual content and see de-biased outcomes with out the necessity for fixed web page refreshes.

The textual content entered by the consumer wanted to be processed by a strong language mannequin to scrutinize for biases. We selected Flan for its robustness, effectivity, and scalability properties. To make the most of Flan, we used SageMaker JumpStart, as proven within the following screenshot. Amazon SageMaker made it straightforward to deploy and handle a pre-trained occasion of Flan, permitting us to deal with optimizing our prompts and queries moderately than managing the mannequin’s efficiency and infrastructure.

Connecting the Flan mannequin to our front-end utility required a sturdy and safe integration, which was achieved utilizing Lambda and API Gateway. With Lambda, we created a serverless perform that communicates instantly with our SageMaker mannequin. We then used API Gateway to create a safe, scalable, and readily accessible endpoint for our React app to invoke the Lambda perform. When a consumer submitted textual content, the app triggered a collection of API calls to the gateway—first to determine if any bias was current, then, if mandatory, further queries to determine, find, and neutralize the bias. All these requests had been routed via the Lambda perform after which to our SageMaker mannequin.

Our closing job within the improvement course of was the number of prompts to question the language mannequin. Right here, the CrowS-Pairs dataset performed an instrumental position as a result of it supplied us with actual examples of biased textual content, which we utilized to fine-tune our requests. We chosen the prompts by an iterative course of, with the target of maximizing accuracy in bias detection inside this dataset.

Wrapping up the method, we noticed a seamless operational circulate within the completed utility. The method begins with a consumer submitting textual content for evaluation, which is then despatched through a POST request to our safe API Gateway endpoint. This triggers the Lambda perform, which communicates with the SageMaker endpoint. Consequently, the Flan mannequin receives a collection of queries. The primary checks for the presence of any biases within the textual content. If biases are detected, further queries are deployed to find, determine, and neutralize these biased parts. The outcomes are then returned via the identical path—first to the Lambda perform, then via the API Gateway, and finally again to the consumer. If any bias was current within the unique textual content, the consumer receives a complete evaluation indicating the sorts of biases detected, whether or not racial, ethnic, or gender. Particular sections of the textual content the place these biases had been discovered are highlighted, giving customers a transparent view of the adjustments made. Alongside this evaluation, a brand new, de-biased model of their textual content is offered, successfully reworking probably biased enter right into a extra inclusive narrative.

Within the following sections, we element the steps to implement this resolution.

Arrange the React setting

We started by organising our improvement setting for React. For bootstrapping a brand new React utility with minimal configuration, we used create-react-app:

npx create-react-app my-app

Construct the consumer interface

Utilizing React, we designed a easy interface for customers to enter textual content, with a submission button, a reset button, and overlaying shows for presenting the processed outcomes after they’re out there.

Provoke the Flan mannequin on SageMaker

We used SageMaker to create a pre-trained occasion of the Flan language mannequin with an endpoint for real-time inference. The mannequin can be utilized in opposition to any JSON-structured payload like the next:

payload = {
      text_inputs: "text_inputs",
      max_length: <max_length>,
      num_return_sequences: <num_return_sequences>,
      top_k: <top_k>,
      top_p: <top_p>,
      do_sample: <do_sample>,
      num_beams: <num_beams>,
      seed: <seed>,

Create a Lambda perform

We developed a Lambda perform that interacted instantly with our SageMaker endpoint. The perform was designed to obtain a request with the consumer’s textual content, ahead it to the SageMaker endpoint, and return the refined outcomes, as proven within the following code (ENDPOINT_NAME was arrange because the SageMaker occasion endpoint):

import os
import io
import boto3
import json
import csv

# seize setting variables
runtime= boto3.consumer('runtime.sagemaker')

def lambda_handler(occasion, context):
    information = json.masses(json.dumps(occasion))
    payload = json.dumps(information['data']).encode('utf-8')

    query_response = runtime.invoke_endpoint(

    response_dict = json.masses(query_response['Body'].learn())

    return response_dict['generated_texts']

Arrange API Gateway

We configured a brand new REST API in API Gateway and linked it to our Lambda perform. This connection allowed our React utility to make HTTP requests to the API Gateway, which subsequently triggered the Lambda perform.

Combine the React app with the API

We up to date the React utility to make a POST request to the API Gateway when the submit button was clicked, with the physique of the request being the consumer’s textual content. The JavaScript code we used to carry out the API name is as follows (REACT_APP_AWS_ENDPOINT corresponds to the API Gateway endpoint sure to the Lambda name):

const makeAWSApiCall = (
  ) => {
    const axiosRequestUrl =
      `${course of.env.REACT_APP_AWS_ENDPOINT}`;
    const requestData = {
      text_inputs: textInputs,
      max_length: maxLength,
      num_return_sequences: numReturnSequences,
      top_k: topK,
      top_p: topP,
      do_sample: doSample,
      num_beams: numBeams,
      seed: 8,

    return axios.submit(axiosRequestUrl, { information: requestData });

Optimize immediate choice

To enhance the accuracy of bias detection, we examined completely different prompts in opposition to the CrowS-Pairs dataset. Via this iterative course of, we selected the prompts that gave us the best accuracy.

Deploy and check the React app on Vercel

After constructing the applying, we deployed it on Vercel to make it publicly accessible. We performed in depth exams to make sure the applying functioned as anticipated, from the consumer interface to the responses from the language mannequin.

These steps laid the groundwork for creating our utility for analyzing and de-biasing textual content. Regardless of the inherent complexity of the method, the usage of instruments like SageMaker, Lambda, and API Gateway streamlined the event, permitting us to deal with the core purpose of the undertaking—figuring out and eliminating biases in textual content.


SageMaker JumpStart presents a handy technique to discover the options and capabilities of SageMaker. It offers curated one-step options, instance notebooks, and deployable pre-trained fashions. These assets will let you rapidly study and perceive SageMaker. Moreover, you’ve got the choice to fine-tune the fashions and deploy them based on your particular wants. Entry to JumpStart is out there via Amazon SageMaker Studio or programmatically utilizing the SageMaker APIs.

On this submit, you realized how a pupil Hackathon workforce developed an answer in a short while utilizing SageMaker JumpStart, which exhibits the potential of AWS and SageMaker JumpStart in enabling speedy improvement and deployment of subtle AI options, even by small groups or people.

To study extra about utilizing SageMaker JumpStart, discuss with Instruction fine-tuning for FLAN T5 XL with Amazon SageMaker Jumpstart and Zero-shot prompting for the Flan-T5 foundation model in Amazon SageMaker JumpStart.

ETH Analytics Membership hosted ‘ETH Datathon,’ an AI/ML hackathon that attracts greater than 150 contributors from ETH Zurich, College of Zurich, and EPFL. The occasion options workshops led by trade leaders, a 24-hour coding problem, and beneficial networking alternatives with fellow college students and trade professionals. Nice because of the ETH Hackathon workforce: Daniele Chiappalupi, Athina Nisioti, and Francesco Ignazio Re, in addition to the remainder of AWS organizing workforce: Alice Morano, Demir Catovic, Iana Peix, Jan Oliver Seidenfuss, Lars Nettemann, and Markus Winterholer.

The content material and opinions on this submit are these of the third-party creator and AWS just isn’t liable for the content material or accuracy of this submit.

Concerning the authors

Jun Zhang is a Options Architect based mostly in Zurich. He helps Swiss clients architect cloud-based options to attain their enterprise potential. He has a ardour for sustainability and strives to unravel present sustainability challenges with expertise. He’s additionally an enormous tennis fan and enjoys enjoying board video games quite a bit.

Mohan Gowda leads Machine Studying workforce at AWS Switzerland. He works primarily with Automotive clients to develop revolutionary AI/ML options and platforms for subsequent era autos. Earlier than working with AWS, Mohan labored with a International Administration Consulting agency with a deal with Technique & Analytics. His ardour lies in linked autos and autonomous driving.

Matthias Egli is the Head of Training in Switzerland. He’s an enthusiastic Crew Lead with a broad expertise in enterprise improvement, gross sales, and advertising and marketing.

Kemeng Zhang is an ML Engineer based mostly in Zurich. She helps world clients design, develop, and scale ML-based functions to empower their digital capabilities to extend enterprise income and cut back value. She can be very obsessed with creating human-centric functions by leveraging data from behavioral science. She likes enjoying water sports activities and strolling canines.

Daniele Chiappalupi is a current graduate from ETH Zürich. He enjoys each facet of software program engineering, from design to implementation, and from deployment to upkeep. He has a deep ardour for AI and eagerly anticipates exploring, using, and contributing to the newest developments within the area. In his free time, he loves going snowboarding throughout colder months and enjoying pick-up basketball when the climate warms up.

Leave a Reply

Your email address will not be published. Required fields are marked *