| --- |
| license: mit |
| task_categories: |
| - other |
| tags: |
| - agent |
| - data-analysis |
| - reinforcement-learning |
| - code-agent |
| - harbor |
| - openenv |
| --- |
| |
| # 📊 Data Agent — Harbor (train) |
|
|
| Teach an agent to *actually do data science*. This is a suite of **5,000 hands-on |
| data-analysis tasks**: each one drops your agent into a sandbox with a real dataset and a |
| question, and asks it to explore the data, compute the answer, and write it down. Every answer is |
| checked **deterministically — no LLM judge, no guesswork**. |
|
|
| It's packaged in [**Harbor**](https://github.com/huggingface/OpenEnv) format, so it runs as a |
| ready-made agentic environment. |
|
|
| ## Where it comes from |
| Built from the [**jupyter-agent dataset**](https://huggingface.co/datasets/jupyter-agent/jupyter-agent-dataset) |
| — real data-science notebooks over Kaggle datasets. We extracted each question–answer pair and |
| then **verified every task**: strong agent models solve it in a live sandbox and must reproduce |
| the gold answer under deterministic grading. Tasks that couldn't be verified cleanly (ambiguous |
| or un-checkable answers) were dropped. So **every task here is known-solvable and unambiguously |
| gradable.** |
|
|
| ## What's inside |
| - **5,000 verified tasks** |
| - **Difficulty** — easy **1,433** · medium **2,845** · hard **722** (`difficulty_tier`; also `difficulty_level` 1–4) |
| - **Answer types** — numeric 2,906 · short-label 1,409 · list 367 · flexible 152 · yes/no 127 · csv-list 39 |
|
|
| ## How a task is laid out |
| ``` |
| tasks/<task_id>/ |
| task.toml # metadata + the question, gold answer, and grading tolerances |
| instruction.md # the prompt the agent sees |
| environment/ # Dockerfile (shared base image) + data-pull hook |
| tests/ # grader.py (deterministic) + test.sh |
| registry.json # index of every task |
| manifest.parquet # the same metadata as a flat table |
| ``` |
| The dataset's CSV/SQLite files are pulled into `/home/user/input/` when the task starts. |
|
|
| ## How grading works |
| The agent writes its final answer to `/workdir/answer.txt`. `grader.py` then scores it through a |
| ladder of deterministic checks — **exact match → numeric tolerance → list/percent normalization → |
| symbolic (math-verify)** — and returns `1.0` (correct) or `0.0`. No network, no model calls. |
|
|
| ## Run it |
| ```bash |
| # see what resolves and how many tasks load |
| openenv harbor info --dataset HuggingEnvs/data-agent-harbor-train |
| |
| # run your agent/model against the suite |
| openenv harbor run --dataset HuggingEnvs/data-agent-harbor-train --model <your-model> |
| ``` |
| Each task gives the agent one shell/code tool, so **any tool-calling model works**, and grading |
| is completely model-agnostic and offline. |
|
|
