Open Dataset Integration Plan

Recommended order

1. FMA small
   • URL: https://github.com/mdeff/fma
   • Why: easiest small realistic music subset for retrieval experiments
2. MTG-Jamendo
   • URL: https://github.com/MTG/mtg-jamendo-dataset
   • Why: larger CC-licensed corpus with scriptable upstream tooling
3. QBSH / humming corpora
   • Why: add after retrieval baseline is stable

Repo strategy

• Keep external dataset ingestion optional
• Convert external tracks into:
  • catalog.json for searchable references
  • query segment manifests for evaluation
• Start with small local subsets before full-corpus scaling