WarungVisionWeights

models

Model weights trained on Kaggle. This folder is empty in git (see .gitignore at the repo root); large files are distributed via Hugging Face Hub, not via commit.

What must be here before running the backend

models/
  detector.pt   YOLOv8 weights fine-tuned on SKU-110K (see notebooks/)
  encoder/      encoder embedding checkpoint folder (config + weights, fine-tuned on RPC)

How they get here

Automatic (default, no manual step): docker compose up fetches both from the Hugging Face Hub repo named by the HF_REPO_ID environment variable in docker-compose.yml, the first time this folder is empty (backend/entrypoint.sh). This satisfies the MVP scope constraint that docker-compose up must run the whole system with no manual step beyond that command (CLAUDE.md section 2). Nothing to do here as a judge running this repo locally, as long as HF_REPO_ID in docker-compose.yml points at a real public repo containing detector.pt at its root and an encoder/ folder alongside it.

Manual (for local development, or if you don't want the auto-download): place detector.pt and the encoder/ folder exactly as shown in the structure above yourself, before running docker compose up. entrypoint.sh checks for both first and skips the network fetch entirely if they're already present, so this keeps working unchanged.

This folder is mounted to /models inside the backend container via a volume mount in docker-compose.yml. Weights are loaded once at startup and never change during runtime (see CLAUDE.md section 6 at the repo root).