Taylor Reese.

I’m a rising junior studying Data Science at UNC Chapel Hill, but most of what I actually do sits closer to ML systems work — building and training models from first principles to understand how they work.

The frame I keep coming back to: using a model isn’t the same as understanding it. Most of my time is spent rebuilding architectures component-by-component, then putting them through real pretraining on real hardware to see what breaks.

Finished pretraining a 751M-parameter Qwen3 reconstruction on UNC’s Longleaf L40S partition — 50,000 steps across 10 chained SLURM submissions, ~11 wall-clock days, on a curated 13B-token corpus I built from six sources. Loss 11.88 → 2.5186 (perplexity ≈ 12.4).

Next chapter: evaluation suite and supervised fine-tuning.

Last updated · June 2026

I got into this through the usual path — PyTorch tutorials, then architectures-from-papers exercises. Around the same time I started using UNC’s Longleaf cluster for a project that became RQwen3, and the gap between “I read about how transformers work” and “I have a SLURM job stuck in the queue at 2am because the GRES string is wrong” clarified what I find interesting about this field.

I’m drawn to pretraining methodology, data curation, and how small-model behavior diverges from scaled-up versions of the same architecture — particularly the work coming out of groups like UNC’s MURGe-Lab.

Before UNC I was residential at NCSSM (grad. 2024). From that stretch I’m also EMT-trained — clinical rotations on an actual ambulance, cert since lapsed — and lifeguard-certified. The clinicals really showed me, first-hand, how real life can get.

The route in wasn’t linear. Most of what I know about training language models I taught myself, on a real cluster, outside any class — and it’s the part of the work I care about most.