The recent rise in the popularity of large language models has spurred the development of extensive code datasets needed to train them. This has left limited code available for collection and use in the downstream investigation of specific behaviors, or evaluation of large language models without suffering from data contamination. To address this problem, we release The Heap, a large multilingual dataset covering 57 programming languages that has been deduplicated with respect to other open datasets of code, enabling researchers to conduct fair evaluations of large language models without significant data cleaning overhead.

Large language models have become increasingly utilized in programming contexts. However, due to the recent emergence of this trend, some aspects have been overlooked. We propose a research approach that investigates the inner mechanics of transformer networks, on a neuron, layer, and output representation level, to understand whether there is a theoretical limitation that prevents large language models from performing optimally in a multilingual setting.We propose to approach the investigation into the theoretical limitations, by addressing open problems in machine learning for the software engineering community. This will contribute to a greater understanding of large language models for programming-related tasks, making the findings more approachable to practitioners, and simply their implementation in future models.