sergiopaniego's picture
sergiopaniego HF Staff
v1.1: remove HF_TOKEN, pin docker_image, 5 gold fixes, eval_verified list
6c3a735 verified
|
Raw History Blame
6.89 kB
---
license: mit
task_categories:
- other
tags:
- smoldataenvs
- rl-environment
- agent
- data-analysis
- reinforcement-learning
- code-agent
- harbor
- openenv
---
<div align="center">
<img src="https://huggingface.co/datasets/FineEnvs/SmolDataEnvs-harbor-eval/resolve/main/banner.png" alt="SmolDataEnvs" width="100%">
# 📊 SmolDataEnvs: Harbor (eval)
[![Collection](https://img.shields.io/badge/%F0%9F%A4%97%20Collection-SmolDataEnvs-FFD21E?style=for-the-badge&labelColor=1a1a1a)](https://huggingface.co/collections/FineEnvs/smoldataenvs)
[![Harbor Visualiser](https://img.shields.io/badge/%F0%9F%A4%97%20Harbor%20Visualiser-Browse%20tasks-FFD21E?style=for-the-badge&labelColor=1a1a1a)](https://huggingface.co/spaces/HuggingFaceH4/harbor-visualiser?dataset=FineEnvs/SmolDataEnvs-harbor-eval)
</div>
> **5.5K+ RL tasks for hill-climbing small models in code and data science.**
<div align="center">
<img src="https://huggingface.co/datasets/FineEnvs/SmolDataEnvs-harbor-eval/resolve/main/curves.gif" alt="Reward and held-out pass@k climbing over 1,119 GRPO steps" width="100%">
<sub>A 2B model on these tasks. Left: what it optimises. Right: 144 held-out tasks it never trains on.<br>
Two runs over the same 5,000 tasks: <b>shuffled</b> against a <b>curriculum</b> ordered easiest to hardest.</sub>
</div>
The validation suite: **144 tasks**, small enough to run every few hundred training steps without
the eval becoming the expensive part of the loop.
Packaged in [Harbor](https://github.com/openenv/OpenEnv) format, so each task is a ready-made
agentic environment: its own container, its own data, its own verifier.
## What's inside
- **144 verified tasks**: quick validation during a run
- **61 reviewed tasks** in `eval_verified.txt`
- **Difficulty**: easy **16**, medium **74**, hard **54** (`difficulty_tier`, plus `difficulty_level` 1–5)
## Verified subset
`eval_verified.txt` lists 61 of the 144 tasks that passed a review (gold answer reproduced by an
independent solution, one reading of the question, no unstated model setup, no file copied from a
train task). `registry_eval_verified.json` is the same list as a Harbor registry. The task folders
are the same as in the full suite. The criteria are on the [SmolDataEnvs card](https://huggingface.co/datasets/FineEnvs/SmolDataEnvs#verified-subsets).
To fetch only those tasks:
```python
from huggingface_hub import hf_hub_download, snapshot_download
repo = "FineEnvs/SmolDataEnvs-harbor-eval"
ids = open(hf_hub_download(repo, "eval_verified.txt", repo_type="dataset")).read().split()
snapshot_download(repo, repo_type="dataset", local_dir="smoldataenvs-eval-verified",
allow_patterns=[f"tasks/{i}/*" for i in ids])
```
and pass the local folder as `--dataset`.
## Versions
| Revision | What changed |
|---|---|
| `v1.0` | The original release. |
| `v1.1` | Removes `HF_TOKEN` from every `task.toml`, pins `docker_image` to the image the train suite uses (so HF Sandboxes can start the tasks), fixes 5 wrong gold answers (listed in `CHANGELOG.md`) and adds the `eval_verified` list. |
> [!WARNING]
> Don't run tasks from revisions before v1.1 with a token in your environment. Their `task.toml`
> passes `HF_TOKEN` into the sandbox, where the agent can read it and print it into traces and
> logs. The bucket with the data files is public, so no token is needed.
## Known issues (planned for v2)
The splits were drawn task by task, not by source dataset, so most eval and test tasks use a Kaggle
dataset that some train task also uses (122 of 144 eval, 185 of 250 test). Some questions, mostly
about a model's accuracy or feature importance, depend on an unstated split, seed or model. Details
and the reviewed subsets are on the [SmolDataEnvs card](https://huggingface.co/datasets/FineEnvs/SmolDataEnvs).
## 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
```bash
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
```bash
openenv harbor serve \
--dataset FineEnvs/SmolDataEnvs-harbor-eval \
--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
```bash
openenv harbor rollout \
--dataset FineEnvs/SmolDataEnvs-harbor-eval \
--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](https://huggingface.co/datasets/jupyter-agent/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
| Repo | What it is |
|---|---|
| [`SmolDataEnvs`](https://huggingface.co/datasets/FineEnvs/SmolDataEnvs) | the tasks as plain rows, load it and prompt any model |
| [`SmolDataEnvs-sft`](https://huggingface.co/datasets/FineEnvs/SmolDataEnvs-sft) | 4,677 verified agent trajectories, TRL-ready |
| [`SmolDataEnvs-harbor-train`](https://huggingface.co/datasets/FineEnvs/SmolDataEnvs-harbor-train) | 5,000 tasks as Harbor environments |
| [`SmolDataEnvs-harbor-test`](https://huggingface.co/datasets/FineEnvs/SmolDataEnvs-harbor-test) | 250 held-out, deliberately harder |
| [`SmolDataEnvs-harbor-eval`](https://huggingface.co/datasets/FineEnvs/SmolDataEnvs-harbor-eval) | 144 for quick validation during a run |
## Citation
```bibtex
@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}
}
```