Integration with nemo-gym

environments/README.md

@@ -30,6 +30,7 @@ This folder contains installable example environments that showcase common usage
 - **ToolEnv (native function-calling)**
   - **tool_test**: Validates parallel tool calls and checks exact tool usage via `ToolRubric` + custom reward.
   - **wiki_search**: Multi-tool retrieval (search/view/read) with `ToolEnv`; final judgment combined via `RubricGroup` with a `JudgeRubric`.
+  - **nemo_workplace_assistant**: NeMo Gym resource-server adapter example with dynamic per-row tools, session seeding, and `/verify`-based rewards.
 
 ### Sandboxes
 - **PythonEnv (ipython-style REPL)**

environments/nemo_arc_agi/nemo_arc_agi.py

@@ -0,0 +1,23 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym_env import (
+    NemoGymEnv,
+    _build_dataset,
+    _reward_from_verify,
+)
+
+
+def load_environment(
+    dataset_split: str = "example",
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset(resource_server="arc_agi", dataset_split=dataset_split)
+    rubric = vf.Rubric(funcs=[_reward_from_verify], weights=[1.0])
+    return NemoGymEnv(
+        resource_server="arc_agi",
+        dataset=dataset,
+        rubric=rubric,
+        max_turns=1,
+        **kwargs,
+    )

environments/nemo_arc_agi/pyproject.toml

@@ -0,0 +1,24 @@
+[project]
+name = "nemo-arc-agi"
+description = "NeMo Gym arc_agi resource server (ARC-AGI visual grid pattern matching)"
+tags = ["nemo-gym", "knowledge", "single-turn", "sandbox"]
+version = "0.1.0"
+requires-python = ">=3.10"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym @ https://test-files.pythonhosted.org/packages/c0/58/451a826009a0b206c932e1ebde3dcff2a8b31152c77133fdde7e5f7ccd90/nemo_gym-0.2.9892rc0-py3-none-any.whl",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_arc_agi.py", "pyproject.toml"]
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_code_gen/nemo_code_gen.py

@@ -0,0 +1,30 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym_env import (
+    NemoGymEnv,
+    _build_dataset,
+    _reward_from_verify,
+)
+
+
+def load_environment(
+    dataset_split: str = "example",
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset(resource_server="code_gen", dataset_split=dataset_split)
+    rubric = vf.Rubric(funcs=[_reward_from_verify], weights=[1.0])
+    return NemoGymEnv(
+        resource_server="code_gen",
+        dataset=dataset,
+        rubric=rubric,
+        max_turns=1,
+        config_overrides={
+            "domain": "coding",
+            "num_processes": 8,
+            "unit_test_timeout_secs": 10,
+            "debug": False,
+        },
+        extra_pip_packages=["numpy==2.2.6"],
+        **kwargs,
+    )

environments/nemo_code_gen/pyproject.toml

@@ -0,0 +1,24 @@
+[project]
+name = "nemo-code-gen"
+description = "NeMo Gym code_gen resource server (competitive coding with Ray execution)"
+tags = ["nemo-gym", "coding", "single-turn", "sandbox"]
+version = "0.1.0"
+requires-python = ">=3.10"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym @ https://test-files.pythonhosted.org/packages/c0/58/451a826009a0b206c932e1ebde3dcff2a8b31152c77133fdde7e5f7ccd90/nemo_gym-0.2.9892rc0-py3-none-any.whl",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_code_gen.py", "pyproject.toml"]
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_example_multi_step/nemo_example_multi_step.py

@@ -0,0 +1,26 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym_env import (
+    NemoGymEnv,
+    _build_dataset,
+    _reward_from_verify,
+)
+
+
+def load_environment(
+    dataset_split: str = "example",
+    max_turns: int = 8,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset(
+        resource_server="example_multi_step", dataset_split=dataset_split
+    )
+    rubric = vf.Rubric(funcs=[_reward_from_verify], weights=[1.0])
+    return NemoGymEnv(
+        resource_server="example_multi_step",
+        dataset=dataset,
+        rubric=rubric,
+        max_turns=max_turns,
+        **kwargs,
+    )

environments/nemo_example_multi_step/pyproject.toml

@@ -0,0 +1,24 @@
+[project]
+name = "nemo-example-multi-step"
+description = "NeMo Gym example_multi_step resource server (multi-tool synonym lookup)"
+tags = ["nemo-gym", "agent", "tools", "multi-turn", "sandbox"]
+version = "0.1.0"
+requires-python = ">=3.10"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym @ https://test-files.pythonhosted.org/packages/c0/58/451a826009a0b206c932e1ebde3dcff2a8b31152c77133fdde7e5f7ccd90/nemo_gym-0.2.9892rc0-py3-none-any.whl",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_example_multi_step.py", "pyproject.toml"]
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_example_single_tool_call/nemo_example_single_tool_call.py

@@ -0,0 +1,26 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym_env import (
+    NemoGymEnv,
+    _build_dataset,
+    _reward_from_verify,
+)
+
+
+def load_environment(
+    dataset_split: str = "example",
+    max_turns: int = 4,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset(
+        resource_server="example_single_tool_call", dataset_split=dataset_split
+    )
+    rubric = vf.Rubric(funcs=[_reward_from_verify], weights=[1.0])
+    return NemoGymEnv(
+        resource_server="example_single_tool_call",
+        dataset=dataset,
+        rubric=rubric,
+        max_turns=max_turns,
+        **kwargs,
+    )

environments/nemo_example_single_tool_call/pyproject.toml

@@ -0,0 +1,24 @@
+[project]
+name = "nemo-example-single-tool-call"
+description = "NeMo Gym example_single_tool_call resource server (simple weather tool)"
+tags = ["nemo-gym", "agent", "tools", "sandbox"]
+version = "0.1.0"
+requires-python = ">=3.10"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym @ https://test-files.pythonhosted.org/packages/c0/58/451a826009a0b206c932e1ebde3dcff2a8b31152c77133fdde7e5f7ccd90/nemo_gym-0.2.9892rc0-py3-none-any.whl",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_example_single_tool_call.py", "pyproject.toml"]
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_instruction_following/nemo_instruction_following.py

@@ -0,0 +1,29 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym_env import (
+    NemoGymEnv,
+    _build_dataset,
+    _reward_from_verify,
+)
+
+
+def load_environment(
+    dataset_split: str = "example",
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset(
+        resource_server="instruction_following", dataset_split=dataset_split
+    )
+    rubric = vf.Rubric(funcs=[_reward_from_verify], weights=[1.0])
+    return NemoGymEnv(
+        resource_server="instruction_following",
+        dataset=dataset,
+        rubric=rubric,
+        max_turns=1,
+        extra_pip_packages=[
+            "git+https://github.com/abukharin-nv/verifiable-instructions.git",
+            "https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl",
+        ],
+        **kwargs,
+    )

environments/nemo_instruction_following/pyproject.toml

@@ -0,0 +1,24 @@
+[project]
+name = "nemo-instruction-following"
+description = "NeMo Gym instruction_following resource server (IFEval/IFBench instruction following)"
+tags = ["nemo-gym", "instruction-following", "single-turn", "sandbox"]
+version = "0.1.0"
+requires-python = ">=3.10"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym @ https://test-files.pythonhosted.org/packages/c0/58/451a826009a0b206c932e1ebde3dcff2a8b31152c77133fdde7e5f7ccd90/nemo_gym-0.2.9892rc0-py3-none-any.whl",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_instruction_following.py", "pyproject.toml"]
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_math_advanced_calculations/nemo_math_advanced_calculations.py

@@ -0,0 +1,26 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym_env import (
+    NemoGymEnv,
+    _build_dataset,
+    _reward_from_verify,
+)
+
+
+def load_environment(
+    dataset_split: str = "example",
+    max_turns: int = 8,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset(
+        resource_server="math_advanced_calculations", dataset_split=dataset_split
+    )
+    rubric = vf.Rubric(funcs=[_reward_from_verify], weights=[1.0])
+    return NemoGymEnv(
+        resource_server="math_advanced_calculations",
+        dataset=dataset,
+        rubric=rubric,
+        max_turns=max_turns,
+        **kwargs,
+    )

environments/nemo_math_advanced_calculations/pyproject.toml

@@ -0,0 +1,24 @@
+[project]
+name = "nemo-math-advanced-calculations"
+description = "NeMo Gym math_advanced_calculations resource server (counter-intuitive calculator tools)"
+tags = ["nemo-gym", "agent", "multi-turn", "sandbox"]
+version = "0.1.0"
+requires-python = ">=3.10"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym @ https://test-files.pythonhosted.org/packages/c0/58/451a826009a0b206c932e1ebde3dcff2a8b31152c77133fdde7e5f7ccd90/nemo_gym-0.2.9892rc0-py3-none-any.whl",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_math_advanced_calculations.py", "pyproject.toml"]
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1

environments/nemo_math_with_code/nemo_math_with_code.py

@@ -0,0 +1,28 @@
+from typing import Any
+
+import verifiers as vf
+from verifiers.envs.integrations.nemo_gym_env import (
+    NemoGymEnv,
+    _build_dataset,
+    _reward_from_verify,
+)
+
+
+def load_environment(
+    dataset_split: str = "example",
+    max_turns: int = 8,
+    **kwargs: Any,
+) -> vf.Environment:
+    dataset, _ = _build_dataset(
+        resource_server="math_with_code", dataset_split=dataset_split
+    )
+    rubric = vf.Rubric(funcs=[_reward_from_verify], weights=[1.0])
+    return NemoGymEnv(
+        resource_server="math_with_code",
+        dataset=dataset,
+        rubric=rubric,
+        max_turns=max_turns,
+        config_overrides={"max_execution_time": 10},
+        extra_pip_packages=["numpy", "scipy", "pandas"],
+        **kwargs,
+    )

environments/nemo_math_with_code/pyproject.toml

@@ -0,0 +1,24 @@
+[project]
+name = "nemo-math-with-code"
+description = "NeMo Gym math_with_code resource server (math with Python code execution)"
+tags = ["nemo-gym", "math", "multi-turn", "sandbox"]
+version = "0.1.0"
+requires-python = ">=3.10"
+dependencies = [
+    "verifiers>=0.1.11.dev1",
+    "nemo-gym @ https://test-files.pythonhosted.org/packages/c0/58/451a826009a0b206c932e1ebde3dcff2a8b31152c77133fdde7e5f7ccd90/nemo_gym-0.2.9892rc0-py3-none-any.whl",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build]
+include = ["nemo_math_with_code.py", "pyproject.toml"]
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.verifiers.eval]
+num_examples = 5
+rollouts_per_example = 1