5.5K+ RL tasks for hill-climbing small models in code and data science.

The training suite: 5,000 hands-on data-analysis tasks. Each one drops an 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.

Packaged in Harbor format, so each task is a ready-made agentic environment: its own container, its own data, its own verifier.

What's inside

• 5,000 verified tasks: the RL training set
• Difficulty: easy 1,433 · medium 2,845 · hard 722 (difficulty_tier, plus difficulty_level 1–5)

How a task is laid out

tasks/<task_id>/
  task.toml         # metadata, the question, the gold answer, grading tolerances
  instruction.md    # the prompt the agent sees
  environment/      # Dockerfile (shared base image) + the data-pull hook
  tests/            # grader.py (deterministic) + test.sh
registry.json       # the suite manifest
manifest.parquet    # one row per task, for filtering without walking the tree

Install

pip install "openenv[harbor]"   # tested with openenv 0.6.0, needs Python 3.12+

The [harbor] extra is what brings in the sandboxes. A plain pip install openenv gives you the CLI but nothing to run a task in. --sandbox e2b also needs E2B_API_KEY set.

Serve it

openenv harbor serve \
  --dataset FineEnvs/SmolDataEnvs-harbor-train \
  --llm-url http://127.0.0.1:8000/v1 --model <your-model> \
  --port 8000 --capture-port 8100

Pass several with --dataset a,b and each arrives as its own split, which is how you train against -train and validate against -eval from one server.

One rollout, no trainer

openenv harbor rollout \
  --dataset FineEnvs/SmolDataEnvs-harbor-train \
  --llm-url http://127.0.0.1:8000/v1 --model <your-model> \
  --harness opencode --sandbox e2b --task-index 0

Where it comes from

Built from the jupyter-agent dataset, real data-science notebooks over 471 Kaggle datasets. Every question–answer pair was extracted and then verified: strong agent models had to solve the task in a live sandbox and reproduce the gold answer under deterministic grading. Anything ambiguous or un-checkable was dropped. So every task here is known-solvable and unambiguously gradable.

Verified by a checker, not judged by a model. Grading is an exact comparison against a known answer, through a ladder of checks: exact match → numeric with tolerances → list and percent normalisation → symbolic equivalence. No LLM sits in the reward path, so the signal does not drift when you change the grader's model, because there isn't one.

The family

Citation

@misc{fineenvs,
  author = {Kolavi, Adithya S},
  title  = {FineEnvs: Open Source RL Environments for LLM Agents},
  year   = {2026},
  url    = {https://github.com/adithya-s-k/FineEnvs}
}