Jonathan Katzy

Closed-source LLMs hide training data. The Poisoned Chalice competition asks: was this code file in the model’s training set? Build membership-inference methods, compete, and present at FSE/AIWare 2026 (Montréal). razvain.github.io/miacomp/ #LLM4Code #FSE2026 #MembershipInference

📣 Exciting competition for the AI & Software Engineering communities! 🎯 Large language models (LLMs) for code are powerful but they often memorize what they’ve seen. When evaluation datasets overlap with training data, reported performance can be misleading

📄 Paper: ieeexplore.ieee.org/abstract/docum… 📊 Dataset: huggingface.co/datasets/AISE-…

🎯 Impact: Enables reliable LLM performance assessment without the data cleaning overhead. Perfect for researchers who want to focus on evaluation, not data preprocessing! Languages covered: Assembly to Mathematica, C to Haskell, Python to Rust 🌍

🧵Introducing "The Heap" - A contamination-free code dataset for fair LLM evaluation! 🔍 The Problem: 90% of code LLM studies don't deduplicate evaluation data, leading to inflated performance metrics @MalihehIzadi @razvanmp27 @avandeursen Thread 👇

Great insights from @katzy_jonathan! Jonathan Katzy from @serg_delft @AISE_TUDelft just presented "A Qualitative Investigation into LLM-Generated Multilingual Code Comments and Automatic Evaluation Metrics". Exciting work on #LLM #CodeComments #SoftwareEngineering

@katzy_jonathan presenting our #FORGE24 paper on “An Exploratory Investigation into Code License Infringements in Large Language Model Training Datasets”! Talk to him if U are interested in the data used for training a wide range of #llm4code and their implications! @ConfForge

We found that even datasets that claim to contain only code from permissively licensed repositories have a large amount of code also attributed to strong copyleft licenses. Raising the question who should be responsible for adherence to code licenses in public online datasets

We analyzed the prevalence of exact duplicates of code files with strong copyleft licensed code in public code datasets, and searched the comments of code files to identify strong copyleft licenses.

Are foundation models vulnerable to lawsuits regarding the data they used to train? We analyzed the state of LLMs wrt licenses in public code datasets. Checkout the preprint: arxiv.org/abs/2403.15230 and discuss the implications with me at FORGE/@ICSEconf in April. @MalihehIzadi

@ICSEconf @katzy_jonathan @MalihehIzadi resulted in identifying 38M exact duplicates in our strong copyleft dataset. This indicates the pervasive issue of #license inconsistencies in #LLMs4Code. We call upon the community to prioritize developing and adopting best practices for #legal & #ethical dataset creation... 3/n

@ICSEconf @katzy_jonathan @MalihehIzadi 1. "An Exploratory Investigation into Code License Infringements in Large Language Model Training Datasets" 📊 w/ @katzy_jonathan, Razvan Popescu, Arie van Deursen, & @MalihehIzadi Our deep dive into 514 million code files used for training various #LLMs trained on code... 2/n

🎆We're happy to share that 2 full papers from our lab, have been accepted at the International Conference on AI Foundation Models and Software Engineering, co-located with @ICSEconf cc @katzy_jonathan, Razvan, Arie, Tim, Frank, Philippe, Berend, Marc, & @MalihehIzadi #FORGE 1/n

At @AISE_TUDelft, we're happy to wrap up the year with 2 papers accepted at @ICSEconf (ResearchTrack)🎉 -Language Models for Code Completion: A Practical Evaluation -Traces of Memorisation in #LLMs for Code @aalkaswan1, @katzy_jonathan, @timvandamdev, Marc, Razvan & @avandeursen

Heartfelt thanks to Prem @devanbu for spending an enlightening week w @serg_delft & delivering a thought-provoking lecture on quality of code generated by #LLMs in our course (ML4SE'23)! I am sure your expertise added a rich dimension to the learning experience of our students.

Why do #LLMs perform better on some programming tasks than others? In our #IEEESCAM2023 paper, we show early evidence for an inherent difference in representations of tokens between languages. arxiv.org/abs/2308.13354 @MalihehIzadi @avandeursen @ieeescam #ML #LLM #ML4SE