HarborForge

Forge Harbor task directories from any evaluation benchmark.

HarborForge provides the abstract contracts (DataMapper, DatasetHandler) for turning raw benchmark datasets into Harbor-compatible task directories, enabling large-scale parallel agent evaluation.

How it works

Raw benchmark data
      ↓  DataMapper.map()
Harbor task directories
      ↓  harbor jobs start
Agent runs in isolated Docker container
      ↓
Verifier scores the output → reward written to Harbor

Each task directory contains an instruction.md (shown to the agent), a Dockerfile (the agent's environment), and a test.sh verifier that writes a float reward to /logs/verifier/reward.txt.

Installation

pip install harborforge

Or with uv:

uv add harborforge

Usage

Implement DatasetHandler for each dataset type in your benchmark, then DataMapper to iterate over tasks:

from harborforge import DataMapper, DatasetHandler

class MyHandler(DatasetHandler):
    dataset_name = "my_dataset"

    def instruction(self, task_data):
        return f"Solve this: {task_data['problem']}"

    def dockerfile(self, task_data):
        return "FROM python:3.12-slim\nWORKDIR /app\n"

    def test_sh(self, task_data):
        answer = task_data["answer"]
        return f"""#!/bin/bash
mkdir -p /logs/verifier
actual=$(cat /output/answer.txt 2>/dev/null)
[ "$actual" = "{answer}" ] && echo 1 || echo 0 > /logs/verifier/reward.txt
"""

class MyMapper(DataMapper):
    def iter_tasks(self):
        for i, task in enumerate(load_my_benchmark()):
            yield f"my_dataset/{i}", f"my_dataset/{i}", MyHandler(), task

# Generate Harbor task directories
MyMapper().run(output_dir=Path(".data/tasks"), registry_path=Path("registry.json"))

Handler contract

SEPARATE verifier mode activates when any of verifier_dockerfile(), verifier_image_tag(), or verifier_files() is set. Files from verifier_files() are copied into tests/ and the generated Dockerfile adds COPY {dest} /verifier/{dest} lines (top-level directories deduplicated).