Amphibians / Benchmark dataset

Frog choruses: several callers, one recording

Source year 2023 · Note reviewed 2026-10-12

AnuraSet describes 27 hours of expert annotations across 42 Neotropical frog species. The paper reports a 70/30 iterative multilabel split grouped by one-minute recording and a 37.8% overall macro-F1 at threshold 0.5; the rare-species group scored 15.7%. This is a species-identification task, not a call-meaning benchmark.

What we reviewed · Selected evidence reviewed

Full data-availability and split sections and selected results in the paper; official Zenodo API metadata; README, exact pinned MIT license and evaluator source read as text. No replication or code execution.

Not yet independently reviewed by a scientific specialist.

What this does — and doesn’t — establish

The article says the raw data and dataset are CC0, while the versioned Zenodo API record declares CC BY 4.0; reuse rights remain unresolved. The reported split groups 3-second samples by one-minute recording, but does not establish site- or year-held-out generalization. The static evaluator passes sigmoid predictions to BCEWithLogitsLoss for a progress-only loss display; saved metrics are computed separately. Data bytes, weights and outputs were not acquired or reproduced, and no qualified herpetologist has reviewed this note.

What we can build on

Overlapping calls and a long-tailed species distribution make macro metrics and grouping choices visible design decisions. Species detection remains separate from caller intent, signal function and shared meaning.

The next reading task

Keep the dataset metadata-only until the CC0/CC BY conflict is clarified. Before any benchmark, inspect the dataset loader and exact split, obtain qualified herpetology/bioacoustics review, and define site/time generalization and rare-class metrics. Do not execute the reviewed evaluator or acquire the archives as part of this review.

Go to the sources

AnuraSet primary dataset paper ↗
Cañas et al. (2023). Full data-availability and split descriptions and selected baseline results reviewed via PMC; no reproduction.
AnuraSet versioned deposit ↗
Record 8342596; official API declares record-level cc-by and lists five files. Paper says the data are CC0; conflicting declarations remain unresolved and no files were acquired.
AnuraSet reproducible benchmark code ↗
Pinned author-code sample: README, MIT license and evaluator read as text. Evaluator is not executed; the dataset loader and split implementation remain unreviewed.

Original editorial note by Talk2Nature contributors. AI-assisted, source-grounded synthesis; not a peer review or an independent replication. Linked works retain their own rights.