Mind The GeekMind The Geek
University showcase15min
INTERMEDIATE

When LLMs Learn from Bad Code: Assessing the Robustness of models to Low-Quality Code

This study asks whether fine-tuning coding models on an organization’s own codebase can degrade generated code quality, and how quickly. It examines assumptions behind corporate model customization and explains the benchmarks, metrics, and pipeline for engineers, MLOps teams, researchers, and users evaluating AI coding assistants.

talk.summaryAiDisclaimer

IOANNIS DIMITRIADIS
IOANNIS DIMITRIADISErnst and Young

talkDetail.whenAndWhere

Friday, November 20, 16:40-16:55
Room 3 - Alexandros
talks.description
Large Language Models and AI agents have trans-
formed software development to an unprecedented extent, en-
abling new levels of automation, productivity, and developer
assistance. In the context of corporate software development,
companies are increasingly fine-tuning such models on their
proprietary codebases to align generated code with their internal
coding standards, conventions, and development practices. This
practice rests on two assumptions: that the existing organizational
codebase is of sufficiently high quality and that fine-tuning on
this code will not degrade the models’ general performance. The
second assumption gives rise to the central question of this study:
Can fine-tuning a model on an organization’s own codebase
degrade the quality of AI-generated code, and, if so, how rapidly
does this degradation occur?
Target audience: Software engineers and team leads who fine-tune or plan to fine-tune models on internal code, MLOps and platform engineers responsible for training data pipelines, researchers and students working on the intersection of software engineering and machine learning, and anyone evaluating AI coding assistants in an organizational setting. No background in model training is required; the benchmarks, the metrics and the pipeline are explained from the ground up.
degradation
fine-tuning
codebase
softwaredevelopment
talks.speakers
IOANNIS DIMITRIADIS

IOANNIS DIMITRIADIS

Ernst and Young

Greece

I’m a Computer Science student at the University of Macedonia in Thessaloniki, focused on machine learning, and currently an AI & Data intern at Ernst & Young (EY).
I’m fascinated by how models learn from their data. My research sits at the intersection of AI and software engineering. AI coding assistants are quickly becoming part of how software is written, so I’m interested in how reliable they really are and what shapes their behavior. For my thesis, I’m researching how low-quality code in training data affects the performance of fine-tuned LLMs.
Alongside my research, I enjoy building things from scratch. I’m a co-founder of two early-stage ventures, including CrewDeck, a marketplace connecting yacht owners with freelance crew.

Outside of tech, sport has shaped who I am. I played for the U16 and U17 Hellenic national basketball teams, captained PAOK’s U16 and U18 teams, and I’m a black belt (2nd Dan) in Taekwondo and a national champion.
It taught me teamwork, leadership and how to perform under pressure, and I bring that same mindset to my work.