NM i AI 2026 Submission Repo

This repository is the public code submission for the NorgesGruppen Data task in NM i AI 2026.

It contains the source used to train, evaluate, package, and validate our computer-vision submissions. Large datasets, trained weights, generated zips, caches, and local experiment artifacts are intentionally excluded from main.

Review path

Start with SUBMISSION.md. It points to the intended final pipeline and the files reviewers should read first.

The main code areas are:

tasks/cv/                  baseline packaging, validation, and detector tooling
tasks/cv_v2_boxfirst/      RF-DETR proposal generation and fusion runtime
tasks/cv_classifier_push/  classifier and reference-embedder reranking pipeline
tasks/cv_moonshot/         prototype-bank / hybrid CV experiments
tests/                     focused regression tests for the CV runtime helpers

Setup

curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync
cp .env.example .env

Some validation commands add exact runtime dependencies with uv run --with ... so the local environment matches the competition sandbox more closely.

Local validation

Examples used during development:

uv run python tasks/cv_classifier_push/package_rfdetr2xl_stagebe4_quantized_embedder_submission.py
uv run --with onnx==1.17.0 python tasks/cv/check_submission.py --zip submission.zip
uv run --with onnx==1.17.0 --with onnxruntime==1.20.0 --with pycocotools==2.0.7 \
  python tasks/cv/analyze_submission.py --zip submission.zip --images dataset/rf_valid
uv run pytest

Notes

• The final public branch is main.
• Large model artifacts are not committed here; this branch is intentionally source-only.
• Historical and alternative pipelines remain in the repository when they are useful for review, but the primary review path is the one documented in SUBMISSION.md.