Useful Kaggle Links

Competition discussions, notebooks, and resources worth tracking.

Discussions

Link	What	Why It Matters
Faster Perch inference	Optimized Perch embedding extraction	Could significantly cut our wall time bottleneck

Notebooks

Link	What	Score
ProtoSSM v5 Maximum Ensemble	Similar approach to ours (Perch + ProtoSSM). Our pipeline is based on this.	0.925 public LB
Pantanal Distill BirdCLEF2026 Improvement	Knowledge distillation approach for BirdCLEF	0.926 public LB

Datasets

dennyloevlie/birdclef2026-pipeline-inputs

Our combined submission dataset. Only input needed besides the competition data. Contains:

File	Source	Credit
perch_v2.onnx	justinchuby/Perch-onnx	Justin Chuby (ONNX conversion), Google (original Perch v2 model)
labels.csv	google/bird-vocalization-classifier	Google
full_perch_arrays.npz	Computed locally from train soundscapes	Us
full_perch_meta.parquet	Computed locally from train soundscapes	Us
full_oof_meta_features.npz	Computed locally (5-fold OOF)	Us
onnxruntime-*.whl	PyPI	Microsoft (ONNX Runtime)