Imad Azizi

Recent work (Lee et al.) showed something striking: by first training a language model on synthetic, non-linguistic patterns before real text, you can reach the same or better perplexity and convergence with roughly 10x less natural-language data. We build on that idea. I work with Llama models trained from scratch on the Polygl0t LLM Foundry stack across 18 synthetic pattern types (identity, periodic, Dyck, counting languages, neural cellular automata, random), measure each pattern's complexity via gzip, and study how the structure and complexity of the synthetic distribution relate to downstream language-modeling improvement, and how that interacts with model scale.

The previous group's hemophilia-assessment app measured ankle range-of-motion with MediaPipe, and the ankle was its weakest joint. I built an experimental benchmark comparing pose-estimation frameworks against goniometer ground truth and cut ankle error from 12.9° to 5.2° (~59% lower) with RTMPose, bringing markerless ankle measurement into the clinically acceptable range for the first time. The winning framework is being integrated into the app used at UKB Bonn.