An illustrated portrait of a person with dark curly hair and round glasses, in a dark open-collared shirt, looking away to the right, drawn in gold and near-black on cream with a vertical gold band behind them

Tomasz Wierzba

PhD candidate, tutor and bench maintainer. Studies reproducible evaluation pipelines and helps students separate sampling variation from implementation errors.

Role
Tutor
Affiliation
PhD candidate, School of Continuous Improvement
Email
tomasz.wierzba@slop.university
Contact
Ed forum, or before the session while the bench is being re-scored.

A ranking that cannot be reproduced is not ready to support a claim about model quality. Tomasz Wierzba helps students turn that principle into an experiment someone else can run.

Doctoral research

Tomasz is a PhD candidate in Slop University’s School of Continuous Improvement and a doctoral researcher in the Measurement Reliability Group, led by Helen Sandoval. The group studies when an evaluation protocol supports a reliable comparison between generative models.

His research focuses on reproducible evaluation pipelines and numerical stability in Gaussian-distance calculations. His doctoral work asks how to distinguish a ranking change caused by fresh samples from one caused by a changed implementation. Repeating a fixed seed checks repeatability; drawing new seeds investigates sampling variation. Those are different checks.

Teaching and the measurement bench

Tomasz tutors SLOP8412, marks assessments and maintains the course bench. He pins the implementation and re-scores A and B before each session, keeping the conditions explicit so that weekly measurement-log entries can be compared.

In Week 4, that means checking the instrument before trusting its output. In the Python workspace, he helps students isolate one change and retain enough information to repeat it. The workspace’s scalar experiments are practice, not image-based FID or a replacement for the course bench.

Getting useful help

Bring the smallest example that shows the problem, together with the settings and seed. State what you predicted and what actually happened. A screenshot of a score alone cannot distinguish a coding error from sampling variation.

Use the course forum for shared debugging questions, or ask before the session. For research-project and supervision interests, consult Helen’s profile; Tomasz’s role here is doctoral research and course support.