mvxlabs benchmarks

Reproducible, source-backed LLM benchmarks from the mvxlabs.de article series.

Each benchmark is a small, transparent test you can re-run yourself. The philosophy is simple: don't trust a ranking — measure it yourself.

Benchmarks

Test	What it measures	Data sources	License
pattern-rag/	LLM answer quality on source-backed German traffic-data questions (RAG)	BASt (traffic), Copernicus/ERA5 (weather)	Code: MIT · Data: dl-de/by-2-0

Conventions

• One folder per test, named by topic — stable, never renamed (so external links don't break).
• Every test documents its data sources in its own SOURCES.md and in the table above.
• Test runs are dated: runs/YYYY-MM-DD/. A benchmark is always a snapshot in time — models, prices and providers change, so every run carries its real measurement date.
• Code is MIT-licensed (see LICENSE); each data source keeps its own license (see the test's SOURCES.md).

— Marco Fuhrmann · mvxlabs.de