---
license: cc-by-4.0
task_categories:
  - reinforcement-learning
language:
  - en
tags:
  - harbor
  - nemotron-gym
  - rl
  - verifiable-rewards
size_categories:
  - 10K<n<100K
---

# laion/nemotron-gym-reasoning-gym-v2

Harbor-format conversion of [nvidia/Nemotron-RL-ReasoningGym-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-ReasoningGym-v1).

Each row contains:

| column | type | description |
| --- | --- | --- |
| `path` | string | Deterministic short ID (`<family>-<sha256[:12]>.tar.gz`) |
| `task_binary` | binary | Gzipped tar containing the full Harbor task |

The tarball contents follow Harbor's task layout:

```
instruction.md                # Prompt shown to the agent
environment/Dockerfile        # python:3.11-slim-bookworm base + task-specific pip deps
tests/test.sh                 # Verifier entrypoint (writes /logs/verifier/reward.txt)
tests/verifier.py             # Verifier implementation (embedded, deterministic)
tests/verifier_data.json      # Per-task verifier inputs (JSON, no code interpolation)
metadata.json                 # Provenance: source_dataset, row_index, family, ...
task.toml                     # Standard Harbor task config (cpu/memory/timeout defaults)
```

## Conversion details

Generated by the `data/nemotron_gym` adapter in
[OpenThoughts-Agent](https://github.com/open-thoughts/OpenThoughts-Agent).
Conversion is **secure-by-construction**:

- Dataset content is never interpolated into shell, Python, or Dockerfile source.
  All values flow through `tests/verifier_data.json` (JSON, parsed at runtime).
- Base image is name-pinned (`python:3.11-slim-bookworm`); pip specs validated
  against a strict allowlist regex.
- Text fields stripped of C0/C1 control characters; lengths capped; tarball
  paths validated against traversal / NUL / absolute-path attacks.
- Tarballs are deterministic (sorted entries, `mtime=0`, `uid/gid=0`) →
  reproducible bytes.

## Verifier family

`reasoning_gym (delegate to upstream reasoning_gym scorer; normalized-match fallback)` — see the source converter for full details.

## Usage with Harbor

```python
from datasets import load_dataset

ds = load_dataset("laion/nemotron-gym-reasoning-gym-v2", split="train")
print(ds[0]["path"], len(ds[0]["task_binary"]))
```

To run a single task with Harbor:

```bash
# Extract one task to a directory and point Harbor at it
python - <<'PY'
import gzip, io, tarfile
from datasets import load_dataset
ds = load_dataset("laion/nemotron-gym-reasoning-gym-v2", split="train")
row = ds[0]
with tarfile.open(fileobj=io.BytesIO(row["task_binary"]), mode="r:gz") as tar:
    tar.extractall("/tmp/reasoning-gym-task")
PY
harbor run -t /tmp/reasoning-gym-task -e daytona  # or -e docker
```

## Source

This dataset is a derivative of
[nvidia/Nemotron-RL-ReasoningGym-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-ReasoningGym-v1), part of NVIDIA's
[NeMo-Gym collection](https://huggingface.co/collections/nvidia/nemo-gym).


## v2 changelog

- Fixed sandbox build failure: the Dockerfile now installs `build-essential` + `python3-dev` so pip can compile `pycosat` (C extension, sdist-only on PyPI) and `cellpylib` (sdist-only) during `pip install reasoning-gym==0.1.20`. v1 had 0% infra success because the slim-bookworm base image lacked a C toolchain.
- Fixed verifier API mismatch: the embedded verifier now calls `reasoning_gym.get_score_answer_fn(source)` (the actual public API as of v0.1.19+) instead of the non-existent `get_scorer`. The normalized-exact-match fallback is preserved.