FrogNet Research 0.1.0 β€” Perch 2

A provenance-locked binary frog_present specialist head over frozen 1,536-dimensional Google Perch 2 embeddings.

Status: public noncommercial research artifact; this exact checked bundle is privately deployed as a review-only field head. This is not a species classifier, a calibrated probability model, or evidence of ecological occurrence without human review.

This repository is the standardized FrogNet research-family mirror of the checked model bundle and evidence summaries from the tagged GitHub source. The upstream tag and internal model slug retain the original field-probe identity for provenance. The package contains no source audio, coordinates, private observation locations, uploader identities, private manifests, credentials, private paths, or arbitrary Python serialization. Public dataset identifiers and regional benchmark names are retained where needed for provenance.

Release identity

Upstream release: frognet-field-probe-v0.1.0
Model slug:      frognet-dev4-field-probe
Bundle ID:       db54359b42526010a2e7782837d2ff8a5e7d98beeebba9214211a2ee83572fa8
Source commit:   55d42a8c96584e09504b9b61127b0dd9e069cdfc
Source archive:  bc82746ca83b305591073d1aeba8a4598e9860953d9a451d087695ac927d29e1

The Hub package exposes the exact checked JSON/NPZ bundle from that release rather than republishing its outer archive. The runtime verifies the bundle-member checksums and recomputes the bundle ID before scoring. No pickle or joblib loader is required. Renaming the Hub repository does not alter the model bytes, bundle ID, threshold, or deployed inference contract.

Inference contract

Field Value
Input audio to Perch 5 seconds, 32 kHz, mono
Input to FrogNet one 1,536-dimensional Perch 2 embedding
Output uncalibrated ranking score for frog_present
Locked field threshold 0.95
Serialization checked JSON metadata plus NPZ numeric arrays
Perch model-tree SHA-256 3fb2d54b3e34534f1130052b25737e54bbb5ebfd340ec040d4510772b64c81ff

The Perch weights are not included. FrogNet cannot score raw waveforms by itself.

Locked field-probe evidence

At threshold 0.95:

Challenge lane Activations Boundary
Private local frog windows 1,253 / 1,308 narrow fixed-recorder and seasonal coverage
Permissively licensed regional iNaturalist frog recordings 106 / 128 broader taxa and observers, still incomplete
Locked insect, chicken, cat, dog, SmartEars, and laying-hen controls 0 / 3,781 challenge result, not a general field false-positive rate

The highest tested confound score was 0.9216794407 on one dog recording, below the locked threshold. Candidate-training grouped test average precision was 0.9645.

These are challenge-set observations. They do not establish prevalence-adjusted precision, field recall, abundance, species identity, or performance at a new recorder or site.

Operational relationship

In the private field system, each 15-second recording is divided into three non-overlapping five-second windows. Perch produces one embedding per window. InsectNet, ChickenNet, and FrogNet score the same embedding independently, while BirdNET remains a separate whole-recording evidence lane. Every output is retained as a model assertion for review rather than a confirmed biological observation.

The operational repository, raw field audio, sensor details, and private review records are intentionally not published here.

Verify the Hub package

sha256sum --check SHA256SUMS
cd bundle && sha256sum --check SHA256SUMS

The root inventory covers every public payload file except itself, including the nested inventory. The nested inventory is the original runtime bundle inventory from the tagged prerelease.

Included files

  • bundle/model.json β€” checked model metadata and inference contract
  • bundle/weights.npz β€” numeric arrays; load with allow_pickle=False
  • bundle/SHA256SUMS β€” original bundle-member checksums
  • release/frognet-field-probe-v0.1.0.manifest.json β€” release identity and rights lanes
  • release/field_smoke_report.json β€” frozen challenge summaries and embedding hashes
  • release/source-release.json β€” upstream tag, commit, and asset identities
  • SHA256SUMS β€” complete Hub payload inventory

Known limits

  • Most private local positives came from one late-May chorus at one stationary recorder.
  • A strict threshold can miss faint, distant, or mixed frog calls.
  • Regional taxa, observers, devices, weather, and seasons remain incompletely represented.
  • Existing recordings were not automatically backfilled when the field probe was installed.
  • The score is suitable for ranking and review triage, not unattended ecological counts.

Rights and provenance

The training matrix used three rights lanes: private diagnostic material, permissively licensed public material, and research/noncommercial material derived from ESC-50. Because noncommercial source material contributed to training, this prerelease is distributed under CC BY-NC-SA 4.0 for noncommercial research and field evaluation, subject to upstream terms.

No upstream audio is redistributed. This release does not represent the full training corpus as permissively licensed. See the release manifest for the retained hashes and exact public/private boundary.

Source

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Collection including TheVortexProject/frognet-research-0.1.0-perch2