harbor-datasets-mlab

Harbor-format MLAB benchmark (MLAgentBench, arXiv:2310.03302): 9 real ML tasks across tabular / text / vision / graph / segmentation (fathomnet + identify-contrails are ≤5 GB competition subsets, the rest full). A general benchmark for any Harbor agent — the tasks are agent-neutral. registry.json at the root, one folder per task.

Dataset (see registry.json)

• mlab-real — each datasets/mlab-real/<task>/ ships real data (environment/data/), a base instruction.md (task description + data/eval context + GPU hint + submission format — no agent- or library-specific content), a scorer (tests/score.py) and the harbor task config. Any agent reads the same instruction and writes /workspace/submission.csv.

Tasks (-i): spaceship-titanic, house-price, amp-parkinsons, imdb, feedback, cifar10, ogbn-arxiv, fathomnet, identify-contrails.

Run

REPO=https://huggingface.co/datasets/danil-e/harbor-datasets-mlab

# Docker (default):
harbor run --repo $REPO -d mlab-real -i feedback -a <agent> -m <model>

# Docker-less (e.g. running Harbor inside a container) — Singularity/Apptainer,
# daemon-less and GPU-capable:
harbor run --repo $REPO -d mlab-real -i feedback -a <agent> -m <model> -e singularity