harbor-datasets-mlab

Harbor-format MLAgentBench (arXiv:2310.03302), posed under the benchmark's own protocol: every workspace already contains a train.py that runs end to end but scores poorly, and the task is to improve it — not to solve the problem from an empty directory. Task descriptions are MLAgentBench's research_problem.txt verbatim.

10 tasks across tabular / text / vision / graph / segmentation / algorithmic. Agent-neutral and self-contained: registry.json at the root, one folder per task.

What each task contains

datasets/mlab-protocol/<task>/
  environment/Dockerfile      # builds FROM python:3.12-slim — no external base image
  environment/requirements.txt
  environment/prepare.py      # stages the data AND train.py into /workspace at build time
  environment/data/           # the real data + the starter train.py
  instruction.md              # research_problem.txt + submission format (agent-neutral)
  tests/                      # scorer + verifier
  task.toml                   # Harbor task config

The answer key never reaches the workspace — prepare.py puts it aside in /opt/mlab, where the scorer reads it.

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

Data: 8 tasks full; fathomnet and identify-contrails are subsets sized to fit the repository (identify-contrails is 155 records × 3 of 8 frames in float16, so its absolute Dice is not comparable to the competition leaderboard); clrs is synthetic by construction, from the official dm-clrs sampler.

Metrics

Each task is scored by the metric MLAgentBench's own eval.py uses: accuracy (spaceship-titanic, imdb, cifar10, ogbn-arxiv), MAE (house-price), SMAPE (amp-parkinsons), MCRMSE (feedback), MAP@20 (fathomnet), Dice (identify-contrails), pointer accuracy (clrs).

The two arms

Every task ships both prompts, so the comparison is reproducible from this repository alone:

• <task>/instruction.md — the benchmark's own instruction. This is the baseline arm.
• <task>/instruction_c1.md — the same instruction plus the C1 layer: a mandate to use a specialized library for the task's modality, and training-discipline blocks. These are the exact bytes the C1 runs were handed, copied from the runs themselves.

datasets/mlab-protocol/C1_LAYER.md explains the layer and datasets/mlab-protocol/c1_prompt.py is the code that produces it. Any Harbor agent can run either arm by pointing at the corresponding file.

Run

harbor run --local . mlab/house-price