Training a Tiny Token Efficient Trace Analyzer

I never learned higher-level math or how models work in college. With agentic coding tools, I want to see how far I can get with near-zero knowledge. I'm post-training a small open-source model from scratch to learn the internals by doing, then optimizing its inference for SOTA trace analysis. This is my weekly log.

Log

March 14, 2026

Starting to define what “weird” means — building out an initial taxonomy of trace anomalies to label against.