Goblinator | Brainrot Translator

Deployed product in action

sigma sigma on the wall, is this project the skibidiest of them all

Inspiration

From a dating app that matches people by their feet, to an app that purposely tries to anger you, our group spanned many ideas trailing into the absurd. Throughout, we kept one thing in common, the endless stream of incomprehensible "brain rot" which flowed through topics. At one point even we didn't know what we were saying, so we decided, why don’t we make something that can tell me what i’m actually saying?

What it does

Goblinator is a web based translator which takes in brain rot laden text and provides the rough meaning of it in English presented. It works similarly to Google Translate.

How we built it

The main bulk of the initial effort went into developing the AI that was to be used. We knew that most major LLMs, although having alright translations of brain rot, were far from perfect in accuracy. We sought to fine-tune a model that could accurately, and significantly more lightweight, provide translation.

As such we scraped Reddit for sentences and phrases containing brain rot, amassing a dataset of approximately 10000 points. We also generated another 2000 synthetic messages for our data points. We then fine-tuned both a Llama 3.2-3B and a GPT-4o model to better translate brain rot. The frontend of the website is built with React, Vite, and Tailwind, and the backend is built with javascript by constructing APIs to send data to and from servers. The UI is mainly based on Google Translate, and is meant to be simplistic.

Challenges we ran into

The major challenge we ran into was the processing of datasets. Brainrot is inherently nonsensical, which necessitated manual construction of expected outputs. Additionally, Reddit did not provide a clean dataset, which required the removal of an expensive amount of noise, which a significant amount was done by hand. Additionally, there was a significant lack of computing power available for us, short of spending an extensive amount of money, slowing down the process of fine tuning significantly.

Accomplishments that we're proud of

This was the first time most of us worked on AI in any capacity beyond an API call, the fact we were able to fine tune a model locally, to some capacity, under such time constraints, was significant.