TheVortexProject commited on
Commit
a303fc3
·
verified ·
1 Parent(s): 96081fe

Align historical baseline with living Perch 2 deployment

Browse files
Files changed (1) hide show
  1. README.md +39 -7
README.md CHANGED
@@ -17,15 +17,20 @@ datasets:
17
 
18
  InsectNet v0.1.0 is a preserved research classifier for scoring insect- and amphibian-related acoustic classes from frozen BirdNET v2.4 output logits.
19
 
20
- This Hugging Face repository is the model-release mirror. The canonical source, tests, package, and GitHub Release are at [github.com/vortexpjeff/insectnet](https://github.com/vortexpjeff/insectnet).
 
 
 
21
 
22
- This repository intentionally contains one model artifact and one release identity. It does **not** publish live sensor addresses, deployment credentials, exact collection locations, private media, or an unattended capture service.
 
 
23
 
24
  ## Release identity
25
 
26
  ```text
27
  Release: insectnet-v0.1.0
28
- Artifact: classifier.joblib
29
  SHA-256: 5e6ecfc68d78a2cf2e9e9e47da5cb58d696e8de354fd620cfcccc5db9da48702
30
  Bytes: 474,892
31
  Status: historical research reference
@@ -72,10 +77,9 @@ The model file does not embed thresholds, a version number, or its original trai
72
  - device addresses or credentials;
73
  - exact collection locations;
74
  - raw or private field audio;
75
- - later experimental model candidates;
76
  - claims of production readiness.
77
 
78
- The original live sidecar diverged from the public v0.1 source during field experiments. That runtime is not republished here until it can be recovered, tested, and released under a separate reviewed version.
79
 
80
  ## Verify the preserved artifact
81
 
@@ -112,6 +116,34 @@ The command returns one probability per declared class. These scores are model a
112
  - The classifier relies on BirdNET logits and cannot score raw audio by itself.
113
  - There is no validated automated-decision threshold policy in this release.
114
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
115
  ## Provenance
116
 
117
  The surviving records identify these source families:
@@ -121,11 +153,11 @@ The surviving records identify these source families:
121
  - **iNaturalist audio:** licenses vary per recording; the original per-record manifest is unavailable.
122
  - **Private field negatives:** not redistributed.
123
 
124
- See [`PROVENANCE.md`](PROVENANCE.md) and the release manifest for the exact surviving claims and gaps.
125
 
126
  ## Privacy and security
127
 
128
- The public release intentionally omits exact collection location, network topology, account names, credentials, private paths, and raw evidence. See [`SECURITY_AND_PRIVACY.md`](SECURITY_AND_PRIVACY.md).
129
 
130
  ## License
131
 
 
17
 
18
  InsectNet v0.1.0 is a preserved research classifier for scoring insect- and amphibian-related acoustic classes from frozen BirdNET v2.4 output logits.
19
 
20
+ This repository preserves one canonical v0.1 release identity and also carries separately
21
+ versioned Perch 2 research candidates. It does **not** publish live sensor addresses,
22
+ deployment credentials, exact collection locations, private media, or an unattended
23
+ capture service.
24
 
25
+ Public discovery and versioned downloads are organized in the
26
+ [The Vortex Project — Multi-Taxa Bioacoustics](https://huggingface.co/collections/TheVortexProject/the-vortex-project-multi-taxa-bioacoustics-6a6d92a9e0938f507e931388)
27
+ Hugging Face collection. GitHub remains the source, tests, and release-engineering surface.
28
 
29
  ## Release identity
30
 
31
  ```text
32
  Release: insectnet-v0.1.0
33
+ Artifact: src/insectnet/data/classifier.joblib
34
  SHA-256: 5e6ecfc68d78a2cf2e9e9e47da5cb58d696e8de354fd620cfcccc5db9da48702
35
  Bytes: 474,892
36
  Status: historical research reference
 
77
  - device addresses or credentials;
78
  - exact collection locations;
79
  - raw or private field audio;
 
80
  - claims of production readiness.
81
 
82
+ The original v0.1 live sidecar diverged from this public source during early field experiments and is not a current deployment target. It has been superseded operationally by the separately versioned Perch 2 specialist architecture; the preserved v0.1 artifact remains a historical BirdNET-logit reference.
83
 
84
  ## Verify the preserved artifact
85
 
 
116
  - The classifier relies on BirdNET logits and cannot score raw audio by itself.
117
  - There is no validated automated-decision threshold policy in this release.
118
 
119
+ ## Perch 2 research artifacts and field heads
120
+
121
+ Three provenance-locked specialist lines were trained and audited in July 2026. None
122
+ replaces the preserved BirdNET-logit v0.1 artifact. Each line has an explicit public
123
+ research boundary and a separate operational status. FrogNet's public mirror contains
124
+ the exact checked JSON/NPZ bundle used by its private review-only field head; the public
125
+ InsectNet and ChickenNet research artifacts must not be conflated with later private
126
+ field bundles.
127
+
128
+ | Candidate | Status | Key external result |
129
+ |---|---|---|
130
+ | [ChickenNet Research 0.1.0](models/chickennet-research-0.1.0-perch2/MODEL_CARD.md) | frozen public joblib not deployed; later private JSON/NPZ successor operates as a review-only field head | 33/42 broad-head hits on a locked iNaturalist chicken challenge; 10/1,308 candidate activations on a private local confound set |
131
+ | [InsectNet Research 0.2.0](models/insectnet-research-0.2.0-perch2/MODEL_CARD.md) | frozen public joblib not deployed; later private JSON/NPZ successor operates as a review-only field head | 11/26 broad activations on an untouched iNaturalist dog challenge |
132
+ | [FrogNet Research 0.1.0](models/frognet-research-0.1.0-perch2/MODEL_CARD.md) | public noncommercial research artifact; exact strict bundle deployed privately for review | 1,253/1,308 local frog-window activations and 0/3,781 tested confound activations at threshold `0.95` |
133
+
134
+ All three lines consume 1,536-dimensional Google Perch 2 embeddings from five-second,
135
+ 32 kHz mono windows. They include exact model/data hashes, grouped split reports,
136
+ hierarchy contracts, source summaries, and challenge reports. They do not include Perch
137
+ weights or source audio.
138
+
139
+ ## Training strategy
140
+
141
+ [`docs/PERCH2_TRAINING_STRATEGY.md`](docs/PERCH2_TRAINING_STRATEGY.md) records the
142
+ provenance-first design used for the Perch 2 specialist heads. The current FrogNet
143
+ research and operational contract is frozen in [`docs/FROGNET_FIELD_PROBE.md`](docs/FROGNET_FIELD_PROBE.md).
144
+ The dated architecture and artifact-versus-runtime boundary for all three living heads is
145
+ recorded in [`docs/LIVING_DEPLOYMENT.md`](docs/LIVING_DEPLOYMENT.md).
146
+
147
  ## Provenance
148
 
149
  The surviving records identify these source families:
 
153
  - **iNaturalist audio:** licenses vary per recording; the original per-record manifest is unavailable.
154
  - **Private field negatives:** not redistributed.
155
 
156
+ See [`docs/PROVENANCE.md`](docs/PROVENANCE.md) and the release manifest for the exact surviving claims and gaps.
157
 
158
  ## Privacy and security
159
 
160
+ The public release intentionally omits exact collection location, network topology, account names, credentials, private paths, and raw evidence. See [`docs/SECURITY_AND_PRIVACY.md`](docs/SECURITY_AND_PRIVACY.md).
161
 
162
  ## License
163