lepinet-models: the phone-app bundles

This repo is the runtime model store for the lepinet web app. The app fetches these folders directly from here in the browser, so the paths are stable and should not be moved.

Looking for a model to use in your own code? Use one of the general-purpose releases instead. They have full model cards, coherent species/genus/family probabilities and a 10-line onnxruntime example:

• gmougeot/lepinet-bioclip2-vitl14: the most accurate and the best-calibrated (also loads with transformers).
• gmougeot/lepinet-effnetv2s: about 2 points behind it at 1/9 of the size, fast on CPU.
• gmougeot/lepinet-dinov3-convnextl: mid-size (217 M), close behind on accuracy, less reliable confidence; DINOv3 License.

Bundles

folder	model	size	notes
b0-fp16/	EfficientNetV2-B0, distilled from an EfficientNetV2-S teacher	30 MB	the app's default; species macro-F1 0.879 on the GBIF test fold
effnetv2b2-v1/	EfficientNetV2-B2, 256-d bottleneck	54 MB	the original v1 app model, with calibration and thresholds

Each folder is a lepinet-app bundle: config.json (the manifest), model.onnx, taxonomy.json (GBIF keys in logit order, plus parent indices) and names.json, with optional calibration.json and thresholds.json. The input is RGB float32 in [0, 1], NCHW, 256×256, with normalisation baked into the graph. The bundle format is specified in the app's DEVELOPER.md.

Trained on GBIF Lepidoptera images (12,041 species, 4,333 genera, 102 families). The code is at GuillaumeMougeot/lepinet. Not for conservation, pest or toxicity decisions: predictions can be wrong.