BioDCASE 2026 Task 4: BirdSet Dataset

BirdSet_BASEAL contains exported embedding bundles for three BirdSet subsets used in the BASEAL workflow: HSN_BASEAL, POW_BASEAL, and UHH_BASEAL.

Directory Overview

• HSN_BASEAL
• POW_BASEAL
• UHH_BASEAL

Each subdirectory contains a ready-to-load embedding bundle with README.md, labels.csv, metadata.csv, and embeddings/perch_v2/*.npy.

Shared Bundle Format

All three subset bundles share the same export layout and metadata conventions:

• Packaged split scope: train and validation
• Embedding model: Perch v2 (Perch 2.0)
• Embedding dimension: 1536
• Join key across bundle files: filename
• Embedding location: embeddings/perch_v2/*.npy
• Label view: labels.csv:label stores retained processed class names as a semicolon-delimited string; empty cells indicate no retained processed label for that sample
• Metadata scope: metadata.csv stores per-sample provenance, timing, and recording-context fields

Detailed column descriptions, class inventories, and loading examples remain in each subset-specific README.md.

Embedding Model

The exported .npy files in embeddings/perch_v2 were generated with the Perch v2 model handle used for this export:

 https://www.kaggle.com/models/google/bird-vocalization-classifier/tensorFlow2/perch_v2/2.

For model attribution, reference the Perch 2.0 paper:

https://research.google/pubs/perch-20-the-bittern-lesson-for-bioacoustics/.

The official project repository is 

https://github.com/google-research/perch,

which is published under the Apache-2.0 license.

Perch model/software attribution is separate from the BirdSet and source-recording provenance documented in each subset’s metadata.csv.

Subset Summary