# 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** — tasks are agent-neutral and **self-contained**.
`registry.json` at the root, **one folder per task**.

## What each task contains
```
datasets/mlab-real/<task>/
  environment/Dockerfile      # builds FROM python:3.12-slim — no external base image
  environment/requirements.txt
  environment/prepare.py      # stages the data into /workspace at build time
  environment/data/           # the real data
  instruction.md              # task description + submission format (agent-neutral)
  tests/                      # scorer + verifier
  task.toml                   # Harbor task config
```
Nothing outside this repository is required: no private images, no extra repos,
no tokens. `docker build` pulls only `python:3.12-slim` and public PyPI/apt
packages. The agent is installed by Harbor itself, so any agent works.

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

## Run
```bash
REPO=https://huggingface.co/datasets/danil-e/harbor-datasets-mlab
harbor run --repo $REPO -d mlab-real -i feedback -a <agent> -m <model>
```
The task image is built on first run (installs the task's Python stack, so the
first build takes a few minutes; later runs reuse the cached image).

Harbor's other environment backends (`-e ...`, e.g. for hosts without a Docker
daemon) work the same way — see the Harbor documentation for their requirements.