linnaeus

name: "linnaeus" description: "Public linnaeus repo knowledge for model architecture, training/inference entry points, and profiling workflows. Inject before editing polli-labs/linnaeus or integrating with typus taxonomy surfaces."

Public Linnaeus repository knowledge for taxonomy-aware biodiversity modeling.

Quick facts

Repo: polli-labs/linnaeus (public)
Core model families: mFormerV0, mFormerV1, DINOv3MultiHead
Config system: YACS defaults in linnaeus/config.py + YAML experiment configs
Package manager: uv (prefer over pip)

Core entry points

linnaeus/main.py - training entry point
linnaeus/config.py - default config schema and hierarchy
linnaeus/models/ - model implementations and blocks
linnaeus/h5data/ - dataset + dataloader stack
linnaeus/profiling/cli.py - profiling analysis CLI
linnaeus/tools/profiling/run_profiling_trials.py - profiling runner CLI

Fast workflow

cd ~/repo/linnaeus
uv sync --extra dev --extra profiling --extra cpu

# Inspect CLI surfaces
linnaeus-prof --help
linnaeus-prof-run --help

# Local smoke train
python -m linnaeus.main \
  --cfg configs/experiments/examples/aves_smoke.yaml \
  --opts TRAIN.EPOCHS 1 DEBUG.PROFILER.ENABLED False

Guardrails

Keep public repo configs free of secrets/private infra details.
Place internal experiment/runtime configs in private repo surfaces, not under public configs/.
Prefer reproducible trial definitions with explicit branch + commit pinning.

References

docs/profiling/README.md
docs/profiling/prof-cli.md
docs/profiling/prof-run.md
docs/dev/