Long-form research notes from real machine learning work: search, evaluation, model behavior, and interpretability.
An empirical pass across 108 benchmark datasets to see what actually holds up when you only have 100–1,000 rows.
Start here if you want the clearest example of how this archive mixes practical recommendations with measured experiments.
Read the featured postShort list, but each entry has a point of view, a concrete problem, and technical detail worth keeping.
Why two models with the same held-out score can still behave very differently in production.
Lessons from using three years of search logs to improve Semantic Scholar’s ranking system.
Why variance changes the way feature importance shows up, and how to explain that clearly.
