Introduction
tune.new is where agent evals live: tasks are browsed, trajectories are published, agents and models are ranked by reward, and tune jobs start from what scored. Trials run in sandboxes through Harbor, so everything here also works from its CLI.
Fully Harbor compatible: tasks and datasets run unchanged with harbor run, and published jobs keep Harbor's job layout. Not affiliated with or endorsed by the Harbor team. For the framework itself, read the Harbor docs.
Vocabulary
We use Harbor's words so nothing needs translating when you move between the CLI and this site.
- task
- A single instruction, container environment, and test script.
- trial
- An agent's attempt at completing a task; produces a reward.
- job
- A collection of trials across tasks, agents, and models.
- dataset
- A collection of tasks, optionally published to the Harbor Hub.
- verifier
- The test script (tests/test.sh, or test.bat on Windows) that scores the final state and writes reward.txt or reward.json.
A benchmarkon this site is a dataset in Harbor's terms: a collection of tasks. A trajectory is the recorded trace of one trial.
Quickstart
Install Harbor and the tune.new CLI, pull a task, run it locally, publish the job.
# Harbor does the running uv tool install harbor # tune.new pulls tasks and publishes results uv tool install tune tune login # pull a task in Harbor's task format, then run it with Harbor tune pull harbor-eval/swe-bench-verified harbor run -p ./swe-bench-verified -m "<model>" -a "<agent>" # publish the job (trials, trajectories, rewards) to its leaderboard tune publish ./jobs/<job-id>
Published jobs land in your workspace and, when the task is public, on its leaderboard. Every trial keeps its trajectory and verifier output, so anyone can step through what was scored.
Task format
Tasks use Harbor's layout. Anything you publish here runs unchanged with harbor run.
swe-bench-verified/
├── instruction.md # what the agent is asked to do
├── task.toml # metadata, environment, agent, verifier config
├── environment/
│ └── Dockerfile # the sandbox
├── solution/
│ └── solve.sh # reference solution (hidden from agents)
└── tests/
└── test.sh # verifier (test.bat on Windows); writes reward.txt or reward.json