Instructions to use TheVortexProject/insectnet with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Scikit-learn
How to use TheVortexProject/insectnet with Scikit-learn:
from huggingface_hub import hf_hub_download import joblib model = joblib.load( hf_hub_download("TheVortexProject/insectnet", "sklearn_model.joblib") ) # only load pickle files from sources you trust # read more about it here https://skops.readthedocs.io/en/stable/persistence.html - Notebooks
- Google Colab
- Kaggle
Data and model provenance
Preserved claim
This repository preserves one v0.1 classifier artifact with an independently verified SHA-256. The artifact consumes 6,522-dimensional output logits from the declared BirdNET v2.4 FP16 TFLite backbone.
The exact original per-record training snapshot no longer survives. Therefore this release does not claim to be independently reproducible from raw source media.
Surviving source-family record
| Source family | Surviving license information | Release treatment |
|---|---|---|
| InsectSet459 | Dataset card states CC BY 4.0; source material may also be CC0 | Named as a source family; media not redistributed |
| ESC-50 | CC BY-NC 3.0 | Named as a source family; media not redistributed |
| iNaturalist audio | License varies by individual recording | No blanket license asserted; media not redistributed |
| Private field negatives | Private research evidence | Not redistributed and location omitted |
| BirdNET backbone | Separate upstream model terms apply | Backbone is not redistributed here |
Missing evidence
The following v0.1 records are unavailable:
- exact source-file list and per-record licenses;
- exact normalized window manifest;
- grouped train/evaluation split;
- threshold-selection report;
- complete field-review ledger linked to this exact artifact;
- original environment lockfile.
These gaps are recorded rather than reconstructed from later experimental branches.
Release rule
A future model version must include:
- artifact SHA-256 and byte size;
- feature extractor identity and checksum;
- ordered classes and independent thresholds;
- source/window manifest digest;
- per-record source and license provenance;
- grouped split policy and seed;
- scoped metrics and known false positives;
- code commit and runtime source checksum;
- deployment target and rollback artifact;
- privacy-reviewed public documentation.
Model outputs remain assertions for review. They do not become confirmed observations without human validation.