Theoretical and empirical studies [21,77,9,55,49,5,12]postulate that such flatter regions generalize better than sharper minima, e.g., due to the flat minimizer's robustness against loss function shifts between trainandtestdata,asillustrated inFig.1.

This project addresses a critical pedagogical need: offering students continuous, on-demand academic assistance beyond conventional reception hours. I present a domain-specific Retrieval-Augmented Generation (RAG) system powered by a quantized Mistral-7B Instruct model and deployed as a Telegram bot. The assistant enhances learning by delivering real-time, personalized responses aligned with the "Introduction to Parallel Processing" course materials. GPU acceleration significantly improves inference latency, enabling practical deployment on consumer hardware. This approach demonstrates how consumer GPUs can enable affordable, private, and effective AI tutoring for HPC education.

Scale has become a main ingredient in obtaining strong machine learning models. As a result, understanding a model's scaling properties is key to effectively designing both the right training setup as well as future generations of architectures. In this work, we argue that scale and training research has been needlessly complex due to reliance on the cosine schedule, which prevents training across different lengths for the same model size. We investigate the training behavior of a direct alternative -- constant learning rate and cooldowns -- and find that it scales predictably and reliably similar to cosine. Additionally, we show that stochastic weight averaging yields improved performance along the training trajectory, without additional training costs, across different scales. Importantly, with these findings we demonstrate that scaling experiments can be performed with significantly reduced compute and GPU hours by utilizing fewer but reusable training runs.

Inspired by recent research that recommends starting neural networks training with large learning rates (LRs) to achieve the best generalization, we explore this hypothesis in detail. Our study clarifies the initial LR ranges that provide optimal results for subsequent training with a small LR or weight averaging. We find that these ranges are in fact significantly narrower than generally assumed. We conduct our main experiments in a simplified setup that allows precise control of the learning rate hyperparameter and validate our key findings in a more practical setting.

Knowing the language of an input text/audio is a necessary first step for using almost every NLP tool such as taggers, parsers, or translation systems. Language identification is a well-studied problem, sometimes even considered solved; in reality, due to lack of data and computational challenges, current systems cannot accurately identify most of the world's 7000 languages. To tackle this bottleneck, we first compile a corpus, MCS-350, of 50K multilingual and parallel children's stories in 350+ languages. MCS-350 can serve as a benchmark for language identification of short texts and for 1400+ new translation directions in low-resource Indian and African languages. Second, we propose a novel misprediction-resolution hierarchical model, LIMIt, for language identification that reduces error by 55% (from 0.71 to 0.32) on our compiled children's stories dataset and by 40% (from 0.23 to 0.14) on the FLORES-200 benchmark. Our method can expand language identification coverage into low-resource languages by relying solely on systemic misprediction patterns, bypassing the need to retrain large models from scratch.

Federated Learning (FL) aims to learn a global model from distributed users while protecting their privacy. However, when data are distributed heterogeneously the learning process becomes noisy, unstable, and biased towards the last seen clients' data, slowing down convergence. To address these issues and improve the robustness and generalization capabilities of the global model, we propose WIMA (Window-based Model Averaging). WIMA aggregates global models from different rounds using a window-based approach, effectively capturing knowledge from multiple users and reducing the bias from the last ones. By adopting a windowed view on the rounds, WIMA can be applied from the initial stages of training. Importantly, our method introduces no additional communication or client-side computation overhead. Our experiments demonstrate the robustness of WIMA against distribution shifts and bad client sampling, resulting in smoother and more stable learning trends. Additionally, WIMA can be easily integrated with state-of-the-art algorithms. We extensively evaluate our approach on standard FL benchmarks, demonstrating its effectiveness.