Björn Bebensee Email & Phone Number

London, United Kingdom

Working on pre-training better language models faster

Seoul, South Korea

Training large language models for dialogue applications: Pre-training (65B+ parameter) multi-lingual generative language models for few-shot and in-context learning in dialogue applications. Selected and prepared training data, optimized speed of pre-training pipeline and applied instruction fine-tuning with diverse alignment data.Research on dialogue state tracking and zero-shot intent classification: Proposed and developed a novel span-based approach to dialogue state tracking which achieves state-of-the-art performance in schema-guided dialogue, published at ACL 2023.Bixby NLU pipeline: Development of a LLM-based NLU pipeline for task-oriented dialogue with the Bixby virtual assistant.Samsung Research, Foundation Model Team

Seoul, South Korea

Research on machine learning for vision and language with a special focus on video and visual question answering. Publications on attention for multimodal video QA.Biointelligence Laboratory, advisor: Byoung-Tak Zhang

Aachen, North Rhine-Westphalia, Germany

Participated in a software engineering project-based internship as part of my degree programme. This project was organized as an industry cooperation with ITERGO. Built a web application to track ERGO's Enterprise Architecture IT infrastructure, lead interviews with employees and create quarterly reports.During the software project I have• built back-end functionality in Node.js and with Postgres• built front-end functionality with pug.js and Bootstrap• packaged and shipped the application using Docker• developed using Git and CI/CD• utilized Jira and Confluence

Master'S Degree, Computer Science And Engineering

Thesis: Co-attentional Transformers for Video Story Understanding Studies focused on machine learning.

Bachelor'S Degree, Computer Science

Thesis: Applying Dynamic Bayesian Networks for Automated Modeling in ArchiMate

Exchange Program, Computer Science