---
license: other
license_name: mixed-permissive
license_link: LICENSE.md
tags:
- math
- reinforcement-learning
- rl-environment
- harbor
size_categories:
- 1K<n<10K
---

# MathConstructOptimize-Envs (Harbor)

The 3,577 tasks of [`amphora/MathConstructOptimize-Envs`](https://huggingface.co/datasets/amphora/MathConstructOptimize-Envs) as runnable [Harbor](https://github.com/laude-institute/harbor) environments. Each task is one directory:

```
tasks/<task_id>/
  task.toml              # metadata (subset, family, problem_key, tier, level, source, license, tags,
                         #   direction/baseline/best_known for optimize), limits, network_mode = "no-network"
  instruction.md         # the prompt; the agent writes its answer to /workdir/answer.json
  environment/Dockerfile # python:3.11 slim + pinned numpy/scipy/sympy/networkx/pysat/z3/pulp/ortools
  environment/instance.json
  tests/test.sh          # runs tests/run.py + tests/check.py, writes /logs/verifier/reward.txt
  solution/solve.sh      # oracle: a known valid answer
```

```bash
uv tool install harbor
huggingface-cli download amphora/MathConstructOptimize-Envs-harbor --repo-type dataset --local-dir mcoe
harbor run -p mcoe/tasks -a oracle -n 16          # validate: every task scores > 0
harbor run -p mcoe/tasks -a terminus-2 -m <model> -n 16
```

**Validation:**
- The oracle agent passes on all 3,577 tasks (construct = 1.0, optimize ≥ 0.1), and a no-op agent scores 0.
- An HTTP canary confirmed that `no-network` blocks internet access inside the task container. Research-level best constructions are public, so do not relax this for RL.
- If you run many trials concurrently, raise Docker's default address pools. Each no-network task creates its own networks, and the defaults are exhausted at roughly 30 concurrent trials.

See the main dataset card for the reward contract, tags, sources and the upstream verifier bugs we fixed.