Data Ethics through AI Fairness

Inspiration

After attending the AI Fairness conference, we were inspired by the growing research around creating machine learning models and large language models that are human centered. Most companies are implementing LLMs into their products and we want to create a platform that ensures the pipeline of training models has checks and balances. This platform gives any user of data to better understand their dataset and biases within their datasets before using it to train their model. We are inspired to create machine learning models that benefit everyone, and to ensure that our world is the safest it can be.

What it does

It is a platform that takes tabular data and outputs details about your dataset including it's bias, toxicity level, and if it is missing features from minority groups.

How we built it

First we looked through literature reviews of the current state of AI Fairness. Using python, we inputed inspiration from FeatureForm to group certain features together with any random datasets. Then we used huggingFace API to help us determine its bias level. Afterwards we output the bias levels from the inputed datasets.

Challenges we ran into

• A lot about the current bias in models and AI LLMs
• The current solutions and technology regarding LLMs
• AI programming
• teamwork
• Business modeling and planning

Accomplishments that we're proud of

A whole new program and platform to evaluate bias through datasets and lots of new researches around the data ethics topic.

What we learned

• A lot about the current bias in models and AI LLMs
• The current solutions and technology regarding LLMs
• AI programming
• teamwork
• Business modeling and planning

What's next for AI Fairness

In terms of measuring bias, we are using huggingFace measurement tool. This is still an ongoing and very active area of research. What's next for us is to bring this into the larger research field and investigate how to best represent transparency, and mitigate bias in machine learning models.

Built With

• featureform
• huggingface
• python
• ux