Below is a list of some projects I’m currently working on, or have worked in the past.
Structural Priming
In this research project at the Context Lab (University of Aberdeen), we aim to understand if LLMs behave similarly to humans when comprehending structure in language. We employ interpretability techniques (such as Integrated Gradients) to understand LLM language comprehension behaviour. Similar data is extracted from humans via eye tracking sensors and EEG.
Attribution and Alignment
In this research project, we aim to understand if LLMs exhibit similar repetition patterns to humans when producing and comprehending open-ended and task-oriented dialogue. This project is a continuation of my BSc thesis.
- Presented at CoNLL 2023 (EMNLP) in Singapore
- ArXiv Link
- Poster (A0 PDF)
- Code
- More Details
HOP
This was a research project I did at the University of Aberdeen. HOP is a novel evaluation metric for word embeddings and taxnomoies. The idea is to harvest a phenomenon observed in the hyperlink network of Wikipedia to quantify the abstraction-capturing capacity of certain word representations.
- Detailed Report
- ā³ HOPv2 is under development
DragonChat
Dragon is a Dungeons & Dragons chatbot based on (at the moment) DialoGPT (medium) and trained on CRD3 (Critical Role Dungeons and Dragons Dataset). In this side-project, I train open-source LLMs on Dungeons & Dragons campaign data to create D&D chatbots. The next iteration of the currently released chatbot will feature a larger LLM trained on more data collected from the internet.
- The initial model is on Hugging Face
CookToday
CookToday is an intelligent kitchen assistant app. It allows users to store and import recipes from any website. It also features a recommendation algorithm. The app and corresponding backend were created during a one-year university group project.
Bamboo App
Created under 40 hours at the Junction hackathon, the Bamboo app aims to reduce the dread caused by doomscrolling. Mostly fully functioning Android app with working backend.
FerroTilt
FerroTilt was a research project I finished while studying EE in Budapest. It is a tilt measurement sensor based on a novel architecture. I won numerous STEM competitions with it and got the chance to present it at various student conferences in Pittsburgh, Stockholm and Tallinn.