feat: nemo gym integration

environments/README.md

@@ -49,6 +49,15 @@ This folder contains installable example environments that showcase common usage
   - **opencode_harbor**: Runs the OpenCode CLI agent on Harbor tasks with API interception via Prime Tunnel.
   - **terminus_harbor**: Runs the Terminus agent on Harbor tasks with API interception via Prime Tunnel.
 
+- **NeMo Gym**: examples of using NVIDIA NeMo Gym environments. Currently requires a local editable install of NeMo Gym (to be updated to PyPI). Some examples to get started (not exhaustive):
+  - **nemo-gym-workplace-assistant**: Multi-step tool use workplace assistant
+  - **nemo-gym-reasoning-gym**: Reasoning Gym tasks with a simple agent
+  - **nemo-gym-reasoning-gym-reflection**: Reasoning Gym with a LangGraph reflection agent
+  - **nemo-gym-structured-outputs**: Structured output tasks
+  - **nemo-gym-code-gen**: Code generation tasks
+  - **nemo-gym-mcqa**: Science multiple choice question answering
+  - **nemo-gym-arc-agi**: ARC-AGI 1 and 2
+
 ### Composition
 - **EnvGroup**
   - **math_group**: Groups two `SingleTurnEnv` tasks (GSM8K + Math) into one environment with shared interface.

environments/nemo_gym/nemo_gym_arc_agi/nemo_gym_arc_agi.py

@@ -0,0 +1,20 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym import (
+    NemoGymEnv,
+    _build_dataset,
+    _resolve_gym_config,
+)
+
+
+def load_environment(
+    dataset_path: str | None = None,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset("arc_agi", "example", dataset_path=dataset_path)
+    return NemoGymEnv(
+        gym_configs=[_resolve_gym_config("arc_agi")],
+        dataset=dataset,
+        **kwargs,
+    )

environments/nemo_gym/nemo_gym_arc_agi/pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "nemo-gym-arc-agi"
+description = "NeMo Gym ARC AGI 1 and 2 environment"
+tags = ["nemo-gym", "knowledge", "single-turn"]
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym>=0.2.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_gym_arc_agi.py", "pyproject.toml"]
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_gym/nemo_gym_code_gen/nemo_gym_code_gen.py

@@ -0,0 +1,20 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym import (
+    NemoGymEnv,
+    _build_dataset,
+    _resolve_gym_config,
+)
+
+
+def load_environment(
+    dataset_path: str | None = None,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset("code_gen", "example", dataset_path=dataset_path)
+    return NemoGymEnv(
+        gym_configs=[_resolve_gym_config("code_gen")],
+        dataset=dataset,
+        **kwargs,
+    )

environments/nemo_gym/nemo_gym_code_gen/pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "nemo-gym-code-gen"
+description = "NeMo Gym code generation environment"
+tags = ["nemo-gym"]
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym>=0.2.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_gym_code_gen.py", "pyproject.toml"]
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_gym/nemo_gym_mcqa/nemo_gym_mcqa.py

@@ -0,0 +1,20 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym import (
+    NemoGymEnv,
+    _build_dataset,
+    _resolve_gym_config,
+)
+
+
+def load_environment(
+    dataset_path: str | None = None,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset("mcqa", "example", dataset_path=dataset_path)
+    return NemoGymEnv(
+        gym_configs=[_resolve_gym_config("mcqa")],
+        dataset=dataset,
+        **kwargs,
+    )

environments/nemo_gym/nemo_gym_mcqa/pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "nemo-gym-mcqa"
+description = "NeMo Gym mcqa environment"
+tags = ["nemo-gym"]
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym>=0.2.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_gym_mcqa.py", "pyproject.toml"]
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_gym/nemo_gym_reasoning_gym/nemo_gym_reasoning_gym.py

@@ -0,0 +1,20 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym import (
+    NemoGymEnv,
+    _build_dataset,
+    _resolve_gym_config,
+)
+
+
+def load_environment(
+    dataset_path: str | None = None,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset("reasoning_gym", "example", dataset_path=dataset_path)
+    return NemoGymEnv(
+        gym_configs=[_resolve_gym_config("reasoning_gym")],
+        dataset=dataset,
+        **kwargs,
+    )

environments/nemo_gym/nemo_gym_reasoning_gym/pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "nemo-gym-reasoning-gym"
+description = "NeMo Gym Reasoning Gym simple agent environment"
+tags = ["nemo-gym"]
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym>=0.2.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_gym_reasoning_gym.py", "pyproject.toml"]
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_gym/nemo_gym_reasoning_gym_parallel_thinking/nemo_gym_reasoning_gym_parallel_thinking.py

@@ -0,0 +1,20 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym import (
+    NemoGymEnv,
+    _build_dataset,
+    _resolve_gym_config,
+)
+
+
+def load_environment(
+    dataset_path: str | None = None,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset("reasoning_gym", "example", dataset_path=dataset_path)
+    return NemoGymEnv(
+        gym_configs=[_resolve_gym_config("reasoning_gym", "parallel_thinking_agent")],
+        dataset=dataset,
+        **kwargs,
+    )

environments/nemo_gym/nemo_gym_reasoning_gym_parallel_thinking/pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "nemo-gym-reasoning-gym-parallel-thinking"
+description = "NeMo Gym Reasoning Gym environment with LangGraph parallel thinking agent"
+tags = ["nemo-gym"]
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym>=0.2.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_gym_reasoning_gym_parallel_thinking.py", "pyproject.toml"]
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_gym/nemo_gym_reasoning_gym_reflection/nemo_gym_reasoning_gym_reflection.py

@@ -0,0 +1,20 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym import (
+    NemoGymEnv,
+    _build_dataset,
+    _resolve_gym_config,
+)
+
+
+def load_environment(
+    dataset_path: str | None = None,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset("reasoning_gym", "example", dataset_path=dataset_path)
+    return NemoGymEnv(
+        gym_configs=[_resolve_gym_config("reasoning_gym", "reflection_agent")],
+        dataset=dataset,
+        **kwargs,
+    )

environments/nemo_gym/nemo_gym_reasoning_gym_reflection/pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "nemo-gym-reasoning-gym-reflection"
+description = "NeMo Gym Reasoning Gym environment with LangGraph reflection agent"
+tags = ["nemo-gym"]
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym>=0.2.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_gym_reasoning_gym_reflection.py", "pyproject.toml"]
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_gym/nemo_gym_structured_outputs/nemo_gym_structured_outputs.py

@@ -0,0 +1,24 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym import (
+    NemoGymEnv,
+    _build_dataset,
+    _resolve_gym_config,
+)
+
+
+def load_environment(
+    dataset_path: str | None = None,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset(
+        "structured_outputs", "example", dataset_path=dataset_path
+    )
+    return NemoGymEnv(
+        gym_configs=[
+            _resolve_gym_config("structured_outputs", "structured_outputs_json")
+        ],
+        dataset=dataset,
+        **kwargs,
+    )

environments/nemo_gym/nemo_gym_structured_outputs/pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "nemo-gym-structured-outputs"
+description = "NeMo Gym structured outputs environment"
+tags = ["nemo-gym"]
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym>=0.2.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_gym_structured_outputs.py", "pyproject.toml"]
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_gym/nemo_gym_workplace_assistant/nemo_gym_workplace_assistant.py

@@ -0,0 +1,22 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym import (
+    NemoGymEnv,
+    _build_dataset,
+    _resolve_gym_config,
+)
+
+
+def load_environment(
+    dataset_path: str | None = None,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset(
+        "workplace_assistant", "example", dataset_path=dataset_path
+    )
+    return NemoGymEnv(
+        gym_configs=[_resolve_gym_config("workplace_assistant")],
+        dataset=dataset,
+        **kwargs,
+    )

environments/nemo_gym/nemo_gym_workplace_assistant/pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "nemo-gym-workplace-assistant"
+description = "NeMo Gym workplace assistant environment"
+tags = ["nemo-gym", "agent", "tools", "multi-turn"]
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym>=0.2.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_gym_workplace_assistant.py", "pyproject.toml"]
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_gym/nemo_gym_xlam_fc/nemo_gym_xlam_fc.py

@@ -0,0 +1,20 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym import (
+    NemoGymEnv,
+    _build_dataset,
+    _resolve_gym_config,
+)
+
+
+def load_environment(
+    dataset_path: str | None = None,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset("xlam_fc", "example", dataset_path=dataset_path)
+    return NemoGymEnv(
+        gym_configs=[_resolve_gym_config("xlam_fc")],
+        dataset=dataset,
+        **kwargs,
+    )