diff --git a/packages/nemo_evaluator_sdk/examples/examples.py b/packages/nemo_evaluator_sdk/examples/examples.py
index 6c8a187c3f..90c8a7a8a5 100644
--- a/packages/nemo_evaluator_sdk/examples/examples.py
+++ b/packages/nemo_evaluator_sdk/examples/examples.py
@@ -6,16 +6,11 @@
 from __future__ import annotations
 
 import asyncio
-import gzip
 import json
 import logging
 import os
-import shutil
-import urllib.error
-import urllib.request
 from collections.abc import Sequence
-from pathlib import Path
-from typing import TYPE_CHECKING, Any, Literal
+from typing import TYPE_CHECKING, Any
 
 from nemo_evaluator_sdk.execution.evaluator import Evaluator
 from nemo_evaluator_sdk.execution.values import EvaluationError
@@ -30,13 +25,11 @@
     Model,
     RangeScore,
     RunConfig,
-    RunConfigOnlineModel,
     SecretRef,
 )
 
 if TYPE_CHECKING:
     import numpy as np
-    from nemo_platform import AsyncNeMoPlatform
 
 
 # --- 1. Defining reusable metric configs and custom metrics ---
@@ -46,11 +39,8 @@
     "You are an evaluator. Rate the response's helpfulness from 0-4. "
     'Return only a JSON object with this shape: {"helpfulness": <integer>}.'
 )
-# Local evaluator and local plugin execution resolve this as an environment variable name.
+# Local evaluator execution resolves this as an environment variable name.
 DEFAULT_API_KEY_SECRET = os.getenv("NMP_EVALUATOR_DEFAULT_API_KEY_SECRET", "NVIDIA_API_KEY")
-DEFAULT_WORKSPACE = os.getenv("NMP_EVALUATOR_DEFAULT_WORKSPACE", "default")
-HELPSTEER2_VALIDATION_JSONL_URL = "https://huggingface.co/datasets/nvidia/HelpSteer2/resolve/main/validation.jsonl.gz"
-_EXAMPLES_DIR = Path(__file__).resolve().parent
 
 
 def configure_example_logging() -> None:
@@ -119,45 +109,6 @@ def configure_example_logging() -> None:
     },
 ]
 
-OFFLINE_HELPFULNESS_DATASET = [
-    {
-        "prompt": "What is the capital of France?",
-        "response": "Paris is the capital of France.",
-        "helpfulness": 4,
-    },
-    {
-        "prompt": "How do I make scrambled eggs?",
-        "response": "Eggs.",
-        "helpfulness": 1,
-    },
-]
-
-
-def get_helpsteer2_dataset() -> Path:
-    """Return a local HelpSteer2 validation JSONL path, downloading it when absent."""
-
-    dataset_dir = _EXAMPLES_DIR / "temp" / "helpsteer2-eval"
-    validation_jsonl = dataset_dir / "validation.jsonl"
-    if validation_jsonl.exists():
-        return validation_jsonl
-
-    dataset_dir.mkdir(parents=True, exist_ok=True)
-    validation_jsonl_gz = dataset_dir / "validation.jsonl.gz"
-    try:
-        with (
-            urllib.request.urlopen(HELPSTEER2_VALIDATION_JSONL_URL, timeout=60) as response,
-            validation_jsonl_gz.open("wb") as gz_file,
-        ):
-            shutil.copyfileobj(response, gz_file)
-    except urllib.error.URLError as e:
-        raise RuntimeError(
-            f"Failed to download HelpSteer2 validation dataset from {HELPSTEER2_VALIDATION_JSONL_URL}: {e}"
-        ) from e
-    with gzip.open(validation_jsonl_gz, "rb") as compressed_file, validation_jsonl.open("wb") as output_file:
-        shutil.copyfileobj(compressed_file, output_file)  # pyright: ignore[reportArgumentType]
-    return validation_jsonl
-
-
 ONLINE_CHAT_PROMPT_TEMPLATE = {"messages": [{"role": "user", "content": "{{item.prompt}}"}]}
 
 
@@ -171,44 +122,6 @@ def get_helpsteer2_dataset() -> Path:
 model_with_custom_headers = model.with_default_headers({"X-My-Header": "value"})
 
 
-async def ensure_remote_evaluator_api_key_secret(workspace: str, client: AsyncNeMoPlatform) -> str:
-    """Resolve API key secret name from env and ensure the secret exists on the platform."""
-    from nemo_platform import ConflictError, NotFoundError
-
-    # API service expects lowercase secret name; if we keep it uppercase, the request will fail.
-    secret_name = DEFAULT_API_KEY_SECRET.lower()
-    try:
-        await client.secrets.retrieve(secret_name, workspace=workspace)
-    except NotFoundError:
-        api_key = os.getenv(DEFAULT_API_KEY_SECRET) or os.getenv("NVIDIA_API_KEY") or os.getenv("NVIDIA_BUILD_API_KEY")
-        if api_key is None:
-            raise RuntimeError(
-                f"Remote online evaluation needs a platform secret named '{secret_name}' in workspace "
-                f"'{workspace}'. Set NVIDIA_BUILD_API_KEY or NVIDIA_API_KEY to let this "
-                "example create it, or create it manually with: "
-                f"nemo secrets create {secret_name} --data '<api-key>' --workspace {workspace}"
-            ) from None
-        try:
-            await client.secrets.create(workspace=workspace, name=secret_name, value=api_key)
-            print(f"Secret {workspace}/{secret_name} created")
-        except ConflictError:
-            pass
-    return secret_name
-
-
-async def model_with_valid_secret(
-    *,
-    execution_mode: Literal["local", "remote"],
-    workspace: str,
-    client: AsyncNeMoPlatform,
-) -> Model:
-    """Return a model configured for local or remote NeMo Platform example execution."""
-    if execution_mode == "remote":
-        secret_name = await ensure_remote_evaluator_api_key_secret(workspace, client)
-        return model.model_copy(update={"api_key_secret": SecretRef(root=secret_name)})
-    return model
-
-
 def create_helpfulness_metric(judge_model: Model) -> LLMJudgeMetric:
     """Build a reusable LLM judge metric for helpfulness scoring."""
     return LLMJudgeMetric(
@@ -626,121 +539,6 @@ def run_sync_example() -> None:
     result.print_summary()
 
 
-# --- 3. NeMo Platform evaluator plugin workflows ---
-async def run_nmp_online_metric_example() -> None:
-    """Run one online metric job locally through the evaluator plugin resource."""
-
-    _print_example_separator(run_nmp_online_metric_example.__name__)
-
-    from nemo_evaluator.sdk.standalone_sdk.backend import AsyncNMPBackend
-    from nemo_platform import AsyncNeMoPlatform
-
-    client = AsyncNeMoPlatform(workspace=DEFAULT_WORKSPACE, timeout=30000.0)
-    try:
-        evaluator = Evaluator(client=AsyncNMPBackend(client.evaluator))
-        result = await evaluator.run(
-            metrics=ExactMatchMetric(reference="{{item.reference}}"),
-            target=model,
-            dataset=ONLINE_EXACT_MATCH_DATASET,
-            prompt_template=ONLINE_CHAT_PROMPT_TEMPLATE,
-            config=RunConfigOnlineModel(parallelism=4),
-        )
-    finally:
-        await client.close()
-
-    print("\nCompleted NeMo Platform online evaluator plugin job locally...")
-    result.print_summary()
-
-
-async def run_nmp_llm_judge_example(
-    is_online: bool = False,
-    limit_samples: int = 2,
-    execution_mode: Literal["local", "remote"] = "local",
-) -> None:
-    """Run a helpfulness judge job locally through the evaluator plugin resource."""
-
-    _print_example_separator(
-        run_nmp_llm_judge_example.__name__,
-        is_online=is_online,
-        limit_samples=limit_samples,
-        execution_mode=execution_mode,
-    )
-
-    from nemo_evaluator.sdk.standalone_sdk.backend import AsyncNMPBackend
-    from nemo_platform import AsyncNeMoPlatform
-
-    nemo_client = AsyncNeMoPlatform(workspace=DEFAULT_WORKSPACE, timeout=30000.0)
-    evaluator_plugin_client = nemo_client.evaluator
-    try:
-        run_kwargs: dict[str, Any] = {}
-        model = await model_with_valid_secret(
-            execution_mode=execution_mode,
-            workspace=DEFAULT_WORKSPACE,
-            client=nemo_client,
-        )
-        evaluator = Evaluator(client=AsyncNMPBackend(evaluator_plugin_client, execution_mode=execution_mode))
-        params: RunConfig | RunConfigOnlineModel = RunConfig(limit_samples=limit_samples)
-
-        if is_online:
-            params = RunConfigOnlineModel(parallelism=4, limit_samples=limit_samples)
-            run_kwargs["target"] = model
-            run_kwargs["prompt_template"] = ONLINE_CHAT_PROMPT_TEMPLATE
-
-        result = await evaluator.run(
-            metrics=create_helpfulness_metric(model),
-            dataset=get_helpsteer2_dataset(),
-            config=params,
-            **run_kwargs,
-        )
-        result.print_summary()
-    finally:
-        await nemo_client.close()
-
-    judge_scores, human_scores = extract_helpfulness_scores(
-        result.row_scores,
-        judge_response_index=1 if is_online else 0,
-    )
-    print(f"\nEvaluated: {len(judge_scores)} samples")
-    if len(judge_scores):
-        print(f"judge avg: {judge_scores.mean()}")
-        print(f"human avg: {human_scores.mean()}")
-
-
-async def run_nmp_benchmark_example() -> None:
-    """Run one NeMo Platform online benchmark example with multiple metrics locally."""
-
-    _print_example_separator(run_nmp_benchmark_example.__name__)
-
-    from nemo_evaluator.sdk.standalone_sdk.backend import AsyncNMPBackend
-    from nemo_platform import AsyncNeMoPlatform
-
-    client = AsyncNeMoPlatform(workspace=DEFAULT_WORKSPACE, timeout=30000.0)
-    try:
-        evaluator = Evaluator(client=AsyncNMPBackend(client.evaluator))
-        exact_match = ExactMatchMetric(reference="{{item.reference}}")
-        contains_required_phrase = StringCheckMetric(
-            operation="contains",
-            left_template="{{sample.output_text}}",
-            right_template="{{item.required_phrase}}",
-        )
-
-        print("\nRunning local NeMo Platform online benchmark evaluation...")
-
-        benchmark_result = await evaluator.run(
-            metrics=[exact_match, contains_required_phrase],
-            target=model,
-            dataset=ONLINE_BENCHMARK_DATASET,
-            prompt_template=ONLINE_CHAT_PROMPT_TEMPLATE,
-            config=RunConfigOnlineModel(parallelism=4),
-        )
-    finally:
-        await client.close()
-    benchmark_result.print_summary()
-    print(f"Online benchmark metric keys: {list(benchmark_result.per_metric)}")
-    print(f"Exact match scores: {benchmark_result.metric_result('exact-match').aggregate_scores.scores}")
-    print(f"String check scores: {benchmark_result.metric_result('string-check').aggregate_scores.scores}")
-
-
 async def run_examples() -> None:
     """Execute the example workflows exposed by this module.
 
@@ -757,12 +555,6 @@ async def run_examples() -> None:
     await run_local_benchmark_with_metric_failure_example()
     await run_local_metric_with_template_failure_example()
 
-    ##### NeMo Platform backend examples #####
-    ### !!! `uv sync --group enabled-plugins` before `nemo run services` to enable evaluator plugin !!!
-    await run_nmp_online_metric_example()
-    await run_nmp_llm_judge_example(is_online=True, execution_mode="remote")
-    await run_nmp_benchmark_example()
-
 
 if __name__ == "__main__":
     configure_example_logging()
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/datasets/loader.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/datasets/loader.py
index 0f7235aa8b..8a6bee0083 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/datasets/loader.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/datasets/loader.py
@@ -3,8 +3,6 @@
 
 """Dataset loader module for evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/datasets/loader.py
-
 import gzip
 import io
 import json
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/execution/README.md b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/execution/README.md
index 08d03db451..274fbc0585 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/execution/README.md
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/execution/README.md
@@ -29,19 +29,9 @@ result = await evaluator.run(
 )
 ```
 
-```python
-# Evaluator plugin execution through the plugin-owned SDK.
-from nemo_evaluator.sdk.standalone_sdk.backend import AsyncNMPBackend
-
-client = AsyncNeMoPlatform(workspace="default")
-evaluator = Evaluator(client=AsyncNMPBackend(client.evaluator))
-result = await evaluator.run(metrics=metric, dataset=data)
-```
-
 ## Design Notes
 
 - `Evaluator()` uses `LocalBackend`.
-- `Evaluator(client=...)` accepts an evaluator backend object. For NeMo Platform
-  execution, wrap the mounted evaluator resource in `NMPBackend` or
-  `AsyncNMPBackend` from the evaluator plugin package.
-- `nemo_platform` and `nemo_evaluator` are optional for local SDK evaluation.
+- `Evaluator(client=...)` accepts an evaluator backend object. Platform-specific
+  backend adapters are provided by
+  [`nemo-evaluator-plugin`](../../../../../plugins/nemo-evaluator/README.md).
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/metrics/aggregation.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/metrics/aggregation.py
index a0261250dd..2e214fe3d3 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/metrics/aggregation.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/metrics/aggregation.py
@@ -3,8 +3,6 @@
 
 """Aggregation data structures and computations for metric results."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/metrics/aggregation.py
-
 import math
 from collections import OrderedDict, defaultdict
 from collections.abc import Mapping, Sequence
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/templates.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/templates.py
index 826e669749..2bf5b8b452 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/templates.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/templates.py
@@ -3,8 +3,6 @@
 
 """Template rendering helpers for evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/templates.py
-
 import json
 from typing import Any
 
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/common.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/common.py
index 1930b94cf2..0fe2dcd2d7 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/common.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/common.py
@@ -3,8 +3,6 @@
 
 """Common value types used throughout evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/values/common.py
-
 from enum import Enum
 
 from pydantic import Field, RootModel
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/datasets.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/datasets.py
index 74eb725e9a..ae636c330d 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/datasets.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/datasets.py
@@ -3,8 +3,6 @@
 
 """Dataset-related value types for evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/values/datasets.py
-
 from typing import Any, TypeAlias
 
 import pyarrow as pa
diff --git a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/results.py b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/results.py
index a629958d6d..b69e04a983 100644
--- a/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/results.py
+++ b/packages/nemo_evaluator_sdk/src/nemo_evaluator_sdk/values/results.py
@@ -3,9 +3,6 @@
 
 """Result types for evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/values/results.py
-# Migrated from: services/evaluator/src/nmp/evaluator/app/values/scores.py
-
 from __future__ import annotations
 
 import json
diff --git a/packages/nemo_evaluator_sdk/tests/test_params.py b/packages/nemo_evaluator_sdk/tests/test_params.py
index 5fac180999..8d9c7bce46 100644
--- a/packages/nemo_evaluator_sdk/tests/test_params.py
+++ b/packages/nemo_evaluator_sdk/tests/test_params.py
@@ -1,61 +1,45 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-from typing import Any
-
 import pytest
-from nemo_evaluator_sdk.values.params import (
-    InferenceParams as SDKInferenceParams,
-)
-from nemo_evaluator_sdk.values.params import RunConfigOnlineModel
-from nmp.common.inference import InferenceParams as CommonInferenceParams
-from pydantic import BaseModel, ValidationError
-
-
-def _normalize_field_contract(model: type[BaseModel]) -> dict[str, dict[str, Any]]:
-    return {
-        name: {
-            "annotation": field.annotation,
-            "default": field.default,
-            "default_factory": field.default_factory,
-            "alias": field.alias,
-            "validation_alias": field.validation_alias,
-            "serialization_alias": field.serialization_alias,
-        }
-        for name, field in model.model_fields.items()
-    }
+from nemo_evaluator_sdk.values.params import InferenceParams, RunConfigOnlineModel
+from pydantic import BaseModel, Field, ValidationError
 
 
 class TestInferenceParamsContract:
-    def test_inference_params_contract_matches_nmp_common(self):
-        expected_model_fields = CommonInferenceParams.model_fields
-        expected_model_fields.pop("model")
-        assert list(SDKInferenceParams.model_fields) == list(expected_model_fields)
-        assert _normalize_field_contract(SDKInferenceParams) == _normalize_field_contract(CommonInferenceParams)
-        assert dict(SDKInferenceParams.model_config) == dict(CommonInferenceParams.model_config)
-        assert callable(getattr(SDKInferenceParams, "check_max_tokens", None))
-        assert callable(getattr(CommonInferenceParams, "check_max_tokens", None))
-
-    @pytest.mark.parametrize("model_cls", [SDKInferenceParams, CommonInferenceParams])
-    def test_inference_params_enforces_same_token_limit_invariant(self, model_cls: type[BaseModel]):
+    def test_inference_params_exposes_expected_sdk_contract(self):
+        assert list(InferenceParams.model_fields) == [
+            "temperature",
+            "max_tokens",
+            "max_completion_tokens",
+            "top_p",
+            "stop",
+        ]
+        assert InferenceParams.model_config.get("extra") == "allow"
+        assert callable(getattr(InferenceParams, "check_max_tokens", None))
+
+    def test_inference_params_enforces_token_limit_invariant(self):
         with pytest.raises(ValidationError, match="max_tokens and max_completion_tokens cannot both be configured"):
-            model_cls(max_tokens=1, max_completion_tokens=1)
+            InferenceParams(max_tokens=1, max_completion_tokens=1)
 
 
 class TestRunConfigOnlineModelInferenceCoercion:
-    def test_evaluation_job_params_coerces_nmp_common_inference_params(self):
-        common_params = CommonInferenceParams.model_validate(
-            {
-                "temperature": 0.25,
-                "max_tokens": 12,
-                "extra_body": {"nvext": {"max_thinking_tokens": 64}},
-            }
-        )
-
-        params = RunConfigOnlineModel.model_validate({"inference": common_params})
-
-        assert isinstance(params.inference, SDKInferenceParams)
-        assert params.inference.model_dump(mode="python") == common_params.model_dump(mode="python", exclude={"model"})
+    def test_evaluation_job_params_coerces_compatible_foreign_inference_model(self):
+        class CompatibleInferenceParams(BaseModel):
+            temperature: float = 0.25
+            max_tokens: int = 12
+            extra_body: dict[str, dict[str, int]] = Field(
+                default_factory=lambda: {"nvext": {"max_thinking_tokens": 64}}
+            )
+
+        params = RunConfigOnlineModel.model_validate({"inference": CompatibleInferenceParams()})
+
+        assert isinstance(params.inference, InferenceParams)
+        assert params.inference.model_dump(mode="python", exclude_none=True) == {
+            "temperature": 0.25,
+            "max_tokens": 12,
+            "extra_body": {"nvext": {"max_thinking_tokens": 64}},
+        }
 
     def test_evaluation_job_params_rejects_incompatible_foreign_model(self):
         class ForeignInferenceParams(BaseModel):
diff --git a/packages/nemo_platform/pyproject.toml b/packages/nemo_platform/pyproject.toml
index 8809e1d1ad..112f73c5e3 100644
--- a/packages/nemo_platform/pyproject.toml
+++ b/packages/nemo_platform/pyproject.toml
@@ -330,7 +330,8 @@ nemo-evaluator-plugin = [
   "cloudpickle>=3.1.1",
   "nemo-evaluator-sdk",
   "nemo-platform-plugin",
-  "nmp-evaluator",
+  "nemo-platform-sdk",
+  "nmp-common",
   "pydantic>=2.10.6",
   "typer>=0.20.0",
 ]
diff --git a/packages/nemo_evaluator_sdk/examples/plugin_examples.py b/plugins/nemo-evaluator/examples/plugin_examples.py
similarity index 99%
rename from packages/nemo_evaluator_sdk/examples/plugin_examples.py
rename to plugins/nemo-evaluator/examples/plugin_examples.py
index 5a699e5e69..fd98c92ce5 100644
--- a/packages/nemo_evaluator_sdk/examples/plugin_examples.py
+++ b/plugins/nemo-evaluator/examples/plugin_examples.py
@@ -16,6 +16,7 @@
 from typing import TYPE_CHECKING, Any, cast
 
 from nemo_evaluator.jobs.evaluate import EvaluateSpec
+from nemo_evaluator.sdk import FilesetRef
 from nemo_evaluator.sdk.resources import AsyncEvaluator
 from nemo_evaluator.sdk.resources import Evaluator as SyncEvaluator
 from nemo_evaluator.sdk.types import (
@@ -40,7 +41,6 @@
 from nemo_evaluator_sdk.values.results import EvaluationResult
 from nemo_platform import APIError, AsyncNeMoPlatform, ConflictError, NeMoPlatform, NotFoundError
 from nemo_platform.types.files import HuggingfaceStorageConfigParam
-from nmp.evaluator.app.values import FilesetRef
 
 if TYPE_CHECKING:
     import numpy as np
diff --git a/plugins/nemo-evaluator/pyproject.toml b/plugins/nemo-evaluator/pyproject.toml
index 8c5f9fb49d..b855b8eb83 100644
--- a/plugins/nemo-evaluator/pyproject.toml
+++ b/plugins/nemo-evaluator/pyproject.toml
@@ -8,8 +8,8 @@ dependencies = [
     "cloudpickle>=3.1.1",
     "nemo-evaluator-sdk",
     "nemo-platform-plugin",
-    "nemo-platform",
-    "nmp-evaluator",
+    "nemo-platform-sdk",
+    "nmp-common",
     "pydantic>=2.10.6",
     "typer>=0.20.0",
 ]
@@ -42,8 +42,8 @@ packages = ["src/nemo_evaluator"]
 [tool.uv.sources]
 nemo-evaluator-sdk = { workspace = true }
 nemo-platform-plugin = { workspace = true }
-nemo-platform = { workspace = true }
-nmp-evaluator = { workspace = true }
+nemo-platform-sdk = { workspace = true }
+nmp-common = { workspace = true }
 
 [dependency-groups]
 dev = ["pytest>=8.3.4", "pytest-asyncio>=0.25.3", "ruff>=0.11.8"]
diff --git a/plugins/nemo-evaluator/src/nemo_evaluator/jobs/evaluate.py b/plugins/nemo-evaluator/src/nemo_evaluator/jobs/evaluate.py
index 07fab01dbb..c072ae07b3 100644
--- a/plugins/nemo-evaluator/src/nemo_evaluator/jobs/evaluate.py
+++ b/plugins/nemo-evaluator/src/nemo_evaluator/jobs/evaluate.py
@@ -12,6 +12,7 @@
 
 from nemo_evaluator.jobs.utils import resolve_run_dataset
 from nemo_evaluator.resolvers import PlatformModelResolver
+from nemo_evaluator.sdk.values.filesets import FilesetRef
 from nemo_evaluator.shared.metric_bundles.bundles import MetricBundle, unbundle_metric
 from nemo_evaluator.shared.metric_bundles.cloudpickle import CloudpickleMetricPayload  # noqa: F401
 from nemo_evaluator_sdk import Evaluator
@@ -32,7 +33,6 @@
 from nemo_platform_plugin.job import NemoJob
 from nemo_platform_plugin.job_context import JobContext
 from nemo_platform_plugin.jobs.api_factory import PlatformJobSpec
-from nmp.evaluator.app.values import FilesetRef
 from pydantic import BaseModel, ConfigDict, Field, model_validator
 
 TargetSpec = Model | Agent
diff --git a/plugins/nemo-evaluator/src/nemo_evaluator/jobs/utils.py b/plugins/nemo-evaluator/src/nemo_evaluator/jobs/utils.py
index c17c7eb0c0..0557bf8436 100644
--- a/plugins/nemo-evaluator/src/nemo_evaluator/jobs/utils.py
+++ b/plugins/nemo-evaluator/src/nemo_evaluator/jobs/utils.py
@@ -5,13 +5,153 @@
 
 from __future__ import annotations
 
-from typing import Any
+from collections.abc import Sequence
+from pathlib import Path
+from typing import Any, cast
 
+from nemo_evaluator.sdk.values.filesets import FilesetRef
+from nemo_evaluator_sdk.datasets.loader import is_glob_pattern
 from nemo_evaluator_sdk.execution.metric_execution import run_sync
+from nemo_evaluator_sdk.metrics.types import MetricsUnion
+from nemo_evaluator_sdk.values import DatasetRows
 from nemo_platform import AsyncNeMoPlatform, NeMoPlatform
+from nemo_platform.filesets import FilesetFileSystem, build_fileset_ref, parse_fileset_ref
 from nemo_platform_plugin.job_context import JobContext
-from nmp.evaluator.app.datasets.nmp_datasets.fileset import download_dataset, download_dataset_sync
-from nmp.evaluator.app.values import FilesetRef
+
+
+def _fileset_ref_parts(dataset: FilesetRef) -> tuple[str, str, str]:
+    """Return workspace, fileset name, and optional file selector."""
+    return parse_fileset_ref(dataset.root, workspace_fallback=None)
+
+
+def _base_download_path(destination: str, workspace: str, fileset: str) -> Path:
+    return Path(destination) / workspace / fileset
+
+
+async def dataset_exists(sdk: AsyncNeMoPlatform, dataset: FilesetRef) -> bool:
+    """Return whether a persisted fileset reference resolves through the Files API."""
+    fs = FilesetFileSystem(sdk=sdk)
+    workspace, fileset, path = _fileset_ref_parts(dataset)
+    base_path = build_fileset_ref("", workspace=workspace, fileset=fileset)
+    try:
+        if not path:
+            return await fs._exists(base_path)
+        if is_glob_pattern(path):
+            files_resource = cast(Any, sdk.files)
+            response = await files_resource.list(remote_path=path, fileset=fileset, workspace=workspace)
+            return bool(response.data)
+        return await fs._exists(build_fileset_ref(path, workspace=workspace, fileset=fileset))
+    except FileNotFoundError:
+        return False
+
+
+async def download_dataset(
+    *,
+    sdk: AsyncNeMoPlatform,
+    dataset: FilesetRef,
+    destination: str,
+) -> Path:
+    """Download a persisted fileset reference to local storage."""
+    workspace, fileset, path = _fileset_ref_parts(dataset)
+    base_dest = _base_download_path(destination, workspace, fileset)
+    base_dest.mkdir(parents=True, exist_ok=True)
+    files_resource = cast(Any, sdk.files)
+
+    if not path:
+        await files_resource.download(
+            remote_path="",
+            fileset=fileset,
+            workspace=workspace,
+            local_path=str(base_dest),
+            max_workers=None,
+        )
+        return base_dest
+
+    if is_glob_pattern(path):
+        await files_resource.download(
+            remote_path=path,
+            fileset=fileset,
+            workspace=workspace,
+            local_path=str(base_dest),
+            max_workers=None,
+        )
+        return base_dest
+
+    local_path = base_dest / path
+    await files_resource.download(
+        remote_path=path,
+        fileset=fileset,
+        workspace=workspace,
+        local_path=str(local_path),
+        max_workers=None,
+    )
+    return local_path
+
+
+def download_dataset_sync(
+    *,
+    sdk: NeMoPlatform,
+    dataset: FilesetRef,
+    destination: str,
+) -> Path:
+    """Sync bridge for fileset reference downloads."""
+    workspace, fileset, path = _fileset_ref_parts(dataset)
+    base_dest = _base_download_path(destination, workspace, fileset)
+    base_dest.mkdir(parents=True, exist_ok=True)
+    files_resource = cast(Any, sdk.files)
+
+    if not path:
+        files_resource.download(
+            remote_path="",
+            fileset=fileset,
+            workspace=workspace,
+            local_path=str(base_dest),
+            max_workers=None,
+        )
+        return base_dest
+
+    if is_glob_pattern(path):
+        files_resource.download(
+            remote_path=path,
+            fileset=fileset,
+            workspace=workspace,
+            local_path=str(base_dest),
+            max_workers=None,
+        )
+        return base_dest
+
+    local_path = base_dest / path
+    files_resource.download(
+        remote_path=path,
+        fileset=fileset,
+        workspace=workspace,
+        local_path=str(local_path),
+        max_workers=None,
+    )
+    return local_path
+
+
+def remote_compile_metric(metric: MetricsUnion | Sequence[MetricsUnion]) -> MetricsUnion:
+    """Return the single metric supported by evaluator plugin job compilation."""
+    if isinstance(metric, Sequence):
+        raise NotImplementedError("Remote benchmark compilation is not implemented for inline evaluator plugin specs.")
+    return cast(MetricsUnion, metric)
+
+
+async def resolve_submit_dataset(
+    async_sdk: AsyncNeMoPlatform,
+    dataset: list[dict[str, object]] | FilesetRef,
+) -> tuple[DatasetRows | FilesetRef, FilesetRef | None]:
+    """Resolve an evaluator plugin dataset for remote metric-job submission.
+
+    FilesetRef datasets are validated via the async SDK and passed through;
+    inline rows are wrapped as ``DatasetRows``.
+    """
+    if isinstance(dataset, FilesetRef):
+        if not await dataset_exists(async_sdk, dataset):
+            raise ValueError(f"FilesetRef dataset does not exist: {dataset.root}")
+        return dataset, dataset
+    return DatasetRows(rows=dataset), None
 
 
 def resolve_run_dataset(
diff --git a/plugins/nemo-evaluator/src/nemo_evaluator/sdk/__init__.py b/plugins/nemo-evaluator/src/nemo_evaluator/sdk/__init__.py
index edd78023bc..8ee0b9e2f6 100644
--- a/plugins/nemo-evaluator/src/nemo_evaluator/sdk/__init__.py
+++ b/plugins/nemo-evaluator/src/nemo_evaluator/sdk/__init__.py
@@ -3,16 +3,20 @@
 
 """Public Evaluator plugin SDK surface."""
 
-from nemo_evaluator.sdk.job_resources import AsyncEvaluatorJobResource, EvaluatorJobResource
-from nemo_evaluator.sdk.resources import AsyncEvaluator, Evaluator
+from typing import TYPE_CHECKING
+
 from nemo_evaluator.sdk.types import (
     ExecutionMode,
-    FilesetRef,
     PluginDatasetInput,
     RunConfig,
     RunConfigOnline,
     RunConfigOnlineModel,
 )
+from nemo_evaluator.sdk.values.filesets import FilesetRef
+
+if TYPE_CHECKING:
+    from nemo_evaluator.sdk.job_resources import AsyncEvaluatorJobResource, EvaluatorJobResource
+    from nemo_evaluator.sdk.resources import AsyncEvaluator, Evaluator
 
 __all__ = [
     "AsyncEvaluator",
@@ -26,3 +30,32 @@
     "FilesetRef",
     "PluginDatasetInput",
 ]
+
+
+def __getattr__(name: str) -> object:
+    """Load resource classes only when callers access them.
+
+    Keep ``import nemo_evaluator.sdk`` lightweight for examples and SDK users
+    that only need value types such as ``FilesetRef`` or ``RunConfig``. Eagerly
+    importing ``resources`` pulls in ``sdk._executor``, which imports
+    ``jobs.evaluate`` and the plugin job/metric-bundle runtime. The lightweight
+    path is used by imports such as ``from nemo_evaluator.sdk import FilesetRef``;
+    preserve this lazy boundary so those imports do not initialize job code.
+    """
+    if name == "AsyncEvaluator":
+        from nemo_evaluator.sdk.resources import AsyncEvaluator
+
+        return AsyncEvaluator
+    if name == "Evaluator":
+        from nemo_evaluator.sdk.resources import Evaluator
+
+        return Evaluator
+    if name == "AsyncEvaluatorJobResource":
+        from nemo_evaluator.sdk.job_resources import AsyncEvaluatorJobResource
+
+        return AsyncEvaluatorJobResource
+    if name == "EvaluatorJobResource":
+        from nemo_evaluator.sdk.job_resources import EvaluatorJobResource
+
+        return EvaluatorJobResource
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
diff --git a/plugins/nemo-evaluator/src/nemo_evaluator/sdk/_executor.py b/plugins/nemo-evaluator/src/nemo_evaluator/sdk/_executor.py
index ff192a29ba..e4042028f4 100644
--- a/plugins/nemo-evaluator/src/nemo_evaluator/sdk/_executor.py
+++ b/plugins/nemo-evaluator/src/nemo_evaluator/sdk/_executor.py
@@ -14,6 +14,7 @@
 
 import httpx
 from nemo_evaluator.jobs.evaluate import EvaluateJob, EvaluateSpec
+from nemo_evaluator.jobs.utils import download_dataset, download_dataset_sync
 from nemo_evaluator.sdk import http_utils
 from nemo_evaluator.sdk.fs_utils import EvaluatorLocalRunResult
 from nemo_evaluator.sdk.job_resources import (
@@ -23,6 +24,7 @@
 )
 from nemo_evaluator.sdk.types import PluginDatasetInput
 from nemo_evaluator.sdk.utils import filter_benchmark_result, filter_evaluation_result
+from nemo_evaluator.sdk.values.filesets import FilesetRef
 from nemo_evaluator.shared.metric_bundles.bundles import MetricBundle, MetricBundlePackager, bundle_metric
 from nemo_evaluator_sdk import Evaluator as SDKEvaluator
 from nemo_evaluator_sdk.datasets.loader import prepare_dataset_rows
@@ -40,8 +42,6 @@
 from nemo_evaluator_sdk.values.results import AggregateFieldName, EvaluationResult
 from nemo_platform import AsyncNeMoPlatform, NeMoPlatform
 from nemo_platform_plugin.scheduler import NemoJobScheduler
-from nmp.evaluator.app.datasets.nmp_datasets.fileset import download_dataset, download_dataset_sync
-from nmp.evaluator.app.values import FilesetRef
 
 _DEFAULT_POLL_INTERVAL_SECONDS = 10.0
 _DEFAULT_JOB_TIMEOUT_SECONDS = 3600.0
diff --git a/plugins/nemo-evaluator/src/nemo_evaluator/sdk/types.py b/plugins/nemo-evaluator/src/nemo_evaluator/sdk/types.py
index 4e672dcfc5..768b881bb1 100644
--- a/plugins/nemo-evaluator/src/nemo_evaluator/sdk/types.py
+++ b/plugins/nemo-evaluator/src/nemo_evaluator/sdk/types.py
@@ -8,13 +8,13 @@
 from pathlib import Path
 from typing import Literal, TypeAlias
 
+from nemo_evaluator.sdk.values.filesets import FilesetRef
 from nemo_evaluator_sdk.values import (
     DatasetInput,
     RunConfig,
     RunConfigOnline,
     RunConfigOnlineModel,
 )
-from nmp.evaluator.app.values import FilesetRef
 
 # TODO: remove this type if we decide nemo_evaluator_sdk will not support remote execution.
 ExecutionMode: TypeAlias = Literal["local", "remote"]
@@ -24,6 +24,7 @@
     "RunConfig",
     "RunConfigOnline",
     "RunConfigOnlineModel",
+    "FilesetRef",
     "ExecutionMode",
     "PluginDatasetInput",
 ]
diff --git a/plugins/nemo-evaluator/src/nemo_evaluator/sdk/values/filesets.py b/plugins/nemo-evaluator/src/nemo_evaluator/sdk/values/filesets.py
new file mode 100644
index 0000000000..30256b2262
--- /dev/null
+++ b/plugins/nemo-evaluator/src/nemo_evaluator/sdk/values/filesets.py
@@ -0,0 +1,33 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Fileset value objects for the evaluator plugin SDK."""
+
+from __future__ import annotations
+
+from pydantic import Field, RootModel
+
+
+class FilesetRef(RootModel[str]):
+    """Reference to a Fileset in the platform Files API.
+
+    A reference is a string with format 'workspace/fileset-name' that points to a
+    persisted fileset entity. When used as a dataset source, all files within the
+    fileset will be downloaded to the job container.
+    """
+
+    root: str = Field(description="Reference to a Fileset (format: workspace/fileset-name).")
+
+    def with_fragment(self, fragment: str) -> FilesetRef:
+        """Return a new fileset reference with a file path fragment appended."""
+        normalized_fragment = fragment.lstrip("/")
+        if not normalized_fragment:
+            raise ValueError("FilesetRef fragment cannot be empty.")
+        if "#" in normalized_fragment:
+            raise ValueError("FilesetRef fragment cannot contain '#'.")
+        if "#" in self.root:
+            raise ValueError("FilesetRef already includes a fragment.")
+        return FilesetRef(root=f"{self.root}#{normalized_fragment}")
+
+
+__all__ = ["FilesetRef"]
diff --git a/plugins/nemo-evaluator/src/nemo_evaluator/tasks/evaluate.py b/plugins/nemo-evaluator/src/nemo_evaluator/tasks/evaluate.py
index 8bacacd2a2..ee87134856 100644
--- a/plugins/nemo-evaluator/src/nemo_evaluator/tasks/evaluate.py
+++ b/plugins/nemo-evaluator/src/nemo_evaluator/tasks/evaluate.py
@@ -5,15 +5,40 @@
 
 from __future__ import annotations
 
+import logging
+import signal
+import sys
+from enum import IntEnum
+from types import FrameType
+
 from nemo_evaluator.jobs.evaluate import EvaluateJob
 from nemo_platform_plugin.tasks.dispatcher import run_task
 from nmp.common.sdk_factory import get_task_sdk
 
+logger = logging.getLogger(__name__)
+
+
+class EvaluateTaskExitCode(IntEnum):
+    """Evaluator task setup exit codes."""
+
+    SDK_INITIALIZATION_FAILED = 2
+
+
+def _shutdown_handler(signum: int, frame: FrameType | None) -> None:
+    logger.warning("Received shutdown signal (%d). Exiting.", signum)
+    raise SystemExit(128 + signum)
+
 
 def main() -> int:
-    """Run the evaluator job in a platform-spawned task process."""
-    return run_task(EvaluateJob, sdk=get_task_sdk("evaluator"))
+    """Build the task SDK and dispatch to the evaluator plugin job."""
+    signal.signal(signal.SIGTERM, _shutdown_handler)
+    try:
+        sdk = get_task_sdk("evaluator")
+    except Exception:
+        logger.exception("Failed to build task SDK for evaluator")
+        return EvaluateTaskExitCode.SDK_INITIALIZATION_FAILED
+    return run_task(EvaluateJob, sdk=sdk)
 
 
 if __name__ == "__main__":
-    raise SystemExit(main())
+    sys.exit(main())
diff --git a/plugins/nemo-evaluator/tests/test_evaluate_job.py b/plugins/nemo-evaluator/tests/test_evaluate_job.py
index 0faf03737b..e7b9f5ff93 100644
--- a/plugins/nemo-evaluator/tests/test_evaluate_job.py
+++ b/plugins/nemo-evaluator/tests/test_evaluate_job.py
@@ -22,6 +22,7 @@
     EvaluateSpec,
 )
 from nemo_evaluator.resolvers import PlatformModelResolver, _parse_required_workspace_name
+from nemo_evaluator.sdk.values.filesets import FilesetRef
 from nemo_evaluator.shared.metric_bundles.bundles import (
     MetricBundle,
     MetricBundlePackager,
@@ -31,6 +32,7 @@
     unbundle_metric,
 )
 from nemo_evaluator.shared.metric_bundles.cloudpickle import CloudpickleMetricBundlePackager
+from nemo_evaluator.tasks.evaluate import EvaluateTaskExitCode
 from nemo_evaluator.tasks.evaluate import main as evaluate_task_main
 from nemo_evaluator_sdk.enums import AgentFormat
 from nemo_evaluator_sdk.metrics.exact_match import ExactMatchMetric
@@ -55,7 +57,6 @@
 from nemo_platform_plugin.job_results import LocalJobResults
 from nemo_platform_plugin.jobs.constants import PERSISTENT_JOB_STORAGE_PATH_ENVVAR
 from nemo_platform_plugin.scheduler import NemoJobScheduler
-from nmp.evaluator.app.values import FilesetRef
 from pydantic import BaseModel, ConfigDict
 from pytest_mock import MockerFixture
 from typer.testing import CliRunner
@@ -974,3 +975,16 @@ def test_main_dispatches_evaluate_job_with_task_sdk(self, mocker: MockerFixture)
         assert exit_code == 0
         get_task_sdk.assert_called_once_with("evaluator")
         run_task.assert_called_once_with(EvaluateJob, sdk=sdk)
+
+    def test_main_returns_setup_exit_code_when_task_sdk_fails(self, mocker: MockerFixture) -> None:
+        get_task_sdk = mocker.patch(
+            "nemo_evaluator.tasks.evaluate.get_task_sdk",
+            side_effect=RuntimeError("boom"),
+        )
+        run_task = mocker.patch("nemo_evaluator.tasks.evaluate.run_task")
+
+        exit_code = evaluate_task_main()
+
+        assert exit_code == EvaluateTaskExitCode.SDK_INITIALIZATION_FAILED
+        get_task_sdk.assert_called_once_with("evaluator")
+        run_task.assert_not_called()
diff --git a/plugins/nemo-evaluator/tests/test_sdk.py b/plugins/nemo-evaluator/tests/test_sdk.py
index 93b6e5dbb5..4bc19ba921 100644
--- a/plugins/nemo-evaluator/tests/test_sdk.py
+++ b/plugins/nemo-evaluator/tests/test_sdk.py
@@ -23,6 +23,7 @@
 from nemo_evaluator.sdk.fs_utils import EvaluatorLocalRunResult
 from nemo_evaluator.sdk.job_resources import AsyncEvaluatorJobResource, EvaluatorJobResource
 from nemo_evaluator.sdk.resources import AsyncEvaluator, Evaluator
+from nemo_evaluator.sdk.values.filesets import FilesetRef
 from nemo_evaluator.shared.metric_bundles.bundles import (
     MetricBundle,
     MetricBundlePackager,
@@ -38,7 +39,6 @@
 from nemo_evaluator_sdk.values.results import AggregatedMetricResult, EvaluationResult
 from nemo_platform import AsyncNeMoPlatform, NeMoPlatform
 from nemo_platform_plugin.jobs.schemas import PlatformJobStatus
-from nmp.evaluator.app.values import FilesetRef
 from pydantic import ValidationError
 from pytest_mock import MockerFixture
 
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/datasets/loader.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/datasets/loader.py
index eae87da00f..14409d42a4 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/datasets/loader.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/datasets/loader.py
@@ -3,8 +3,6 @@
 
 """Dataset loader module for evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/datasets/loader.py
-
 import gzip
 import io
 import json
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/metrics/aggregation.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/metrics/aggregation.py
index 90523d7793..d802bf45f7 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/metrics/aggregation.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/metrics/aggregation.py
@@ -3,8 +3,6 @@
 
 """Aggregation data structures and computations for metric results."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/metrics/aggregation.py
-
 import math
 from collections import OrderedDict, defaultdict
 from collections.abc import Mapping, Sequence
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/templates.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/templates.py
index 826e669749..2bf5b8b452 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/templates.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/templates.py
@@ -3,8 +3,6 @@
 
 """Template rendering helpers for evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/templates.py
-
 import json
 from typing import Any
 
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/common.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/common.py
index 1930b94cf2..0fe2dcd2d7 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/common.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/common.py
@@ -3,8 +3,6 @@
 
 """Common value types used throughout evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/values/common.py
-
 from enum import Enum
 
 from pydantic import Field, RootModel
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/datasets.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/datasets.py
index 74eb725e9a..ae636c330d 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/datasets.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/datasets.py
@@ -3,8 +3,6 @@
 
 """Dataset-related value types for evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/values/datasets.py
-
 from typing import Any, TypeAlias
 
 import pyarrow as pa
diff --git a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/results.py b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/results.py
index 47c216a873..3157bc5685 100644
--- a/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/results.py
+++ b/sdk/python/nemo-platform/src/nemo_platform/beta/evaluator/values/results.py
@@ -3,9 +3,6 @@
 
 """Result types for evaluator SDK runtime."""
 
-# Migrated from: services/evaluator/src/nmp/evaluator/app/values/results.py
-# Migrated from: services/evaluator/src/nmp/evaluator/app/values/scores.py
-
 from __future__ import annotations
 
 import json
diff --git a/tools/nemo-platform-sdk-tools/src/nemo_platform_sdk_tools/sdk/vendor/vendor_package.py b/tools/nemo-platform-sdk-tools/src/nemo_platform_sdk_tools/sdk/vendor/vendor_package.py
index 3ca9d23198..a1edd2945e 100644
--- a/tools/nemo-platform-sdk-tools/src/nemo_platform_sdk_tools/sdk/vendor/vendor_package.py
+++ b/tools/nemo-platform-sdk-tools/src/nemo_platform_sdk_tools/sdk/vendor/vendor_package.py
@@ -586,6 +586,21 @@ def _should_copy_optional_dependency_extra(extra_name: str, target_project_name:
     return True
 
 
+def _generated_optional_dependency_groups(optional: tomlkit.items.Table) -> set[str]:
+    """Return optional-dependency groups owned by the bundle generator."""
+    generated: set[str] = set()
+    saw_marker = False
+    for key, item in optional._value.body:
+        if key is None:
+            if isinstance(item, tomlkit.items.Comment) and item.as_string().strip() == GENERATED_BUNDLE_GROUP_COMMENT:
+                saw_marker = True
+            continue
+        if saw_marker:
+            generated.add(key.key)
+        saw_marker = False
+    return generated
+
+
 def _merge_matching_project_optional_dependencies(
     target_project: dict, source_project: dict, patterns: list[str]
 ) -> None:
@@ -707,6 +722,7 @@ def _process_bundle_packages() -> None:
 
         pyproject = tomlkit.loads(member_pyproject_path.read_text(encoding="utf-8"))
         optional = pyproject["project"].setdefault("optional-dependencies", tomlkit.table())
+        generated_optional_groups = _generated_optional_dependency_groups(optional)
 
         for pkg_name, pkg_config in bundle_config.items():
             deps_group = _bundle_deps_group(pkg_name, pkg_config)
@@ -759,7 +775,11 @@ def _process_bundle_packages() -> None:
                 self_ref_deps = [d for d in filtered_deps if d.startswith(self_ref_prefix)]
 
                 existing = list(optional.get(deps_group, []))
-                existing_regular = [d for d in existing if not d.startswith(self_ref_prefix)]
+                existing_regular = (
+                    []
+                    if deps_group in generated_optional_groups
+                    else [d for d in existing if not d.startswith(self_ref_prefix)]
+                )
                 existing_self_refs = [d for d in existing if d.startswith(self_ref_prefix)]
 
                 merged_regular = merge_dependencies(existing_regular, regular_deps)
diff --git a/tools/nemo-platform-sdk-tools/tests/sdk/vendor/test_vendor_package.py b/tools/nemo-platform-sdk-tools/tests/sdk/vendor/test_vendor_package.py
index 1d1285969d..9df3de6a9a 100644
--- a/tools/nemo-platform-sdk-tools/tests/sdk/vendor/test_vendor_package.py
+++ b/tools/nemo-platform-sdk-tools/tests/sdk/vendor/test_vendor_package.py
@@ -323,6 +323,71 @@ def test_process_bundle_packages_keeps_bundled_dependency_names_without_inheriti
     assert not wrapper_updated["project"]["entry-points"]
 
 
+def test_process_bundle_packages_rebuilds_generated_dependency_groups(tmp_path: Path, monkeypatch) -> None:
+    wrapper_path = tmp_path / "packages/nemo_platform"
+    plugin_path = tmp_path / "plugins/nemo-evaluator"
+    runner_path = tmp_path / "packages/nmp_platform_runner"
+    (plugin_path / "src/nemo_evaluator").mkdir(parents=True)
+    (runner_path / "src/nmp/platform_runner").mkdir(parents=True)
+    wrapper_path.mkdir(parents=True)
+
+    (tmp_path / "pyproject.toml").write_text(
+        """
+[tool.uv.workspace]
+members = ["packages/nemo_platform"]
+""".lstrip(),
+        encoding="utf-8",
+    )
+    (wrapper_path / "pyproject.toml").write_text(
+        """
+[project]
+name = "nemo-platform"
+
+[project.optional-dependencies]
+# Generated from [tool.bundle-package]; do not edit by hand.
+nemo-evaluator-plugin = ["nemo-evaluator-sdk", "nmp-evaluator"]
+
+# Generated from [tool.bundle-package]; do not edit by hand.
+services = ["nemo-platform[core-service]", "old-service"]
+
+[tool.bundle-package]
+nemo-evaluator-plugin = { source = "../../plugins/nemo-evaluator/src/nemo_evaluator", module = "nemo_evaluator" }
+nmp-platform-runner = { source = "../../packages/nmp_platform_runner/src/nmp/platform_runner", module = "nmp/platform_runner", deps_group = "services" }
+""".lstrip(),
+        encoding="utf-8",
+    )
+    (plugin_path / "pyproject.toml").write_text(
+        """
+[project]
+name = "nemo-evaluator-plugin"
+dependencies = ["nemo-evaluator-sdk", "nemo-platform-sdk", "nmp-common", "pydantic>=2.10.6"]
+""".lstrip(),
+        encoding="utf-8",
+    )
+    (runner_path / "pyproject.toml").write_text(
+        """
+[project]
+name = "nmp-platform-runner"
+dependencies = ["rich>=14.1.0"]
+""".lstrip(),
+        encoding="utf-8",
+    )
+
+    monkeypatch.setattr(vendor_package, "NMP_ROOT_PATH", tmp_path)
+    vendor_package._process_bundle_packages()
+
+    wrapper_updated = tomlkit.parse((wrapper_path / "pyproject.toml").read_text(encoding="utf-8"))
+    optional = wrapper_updated["project"]["optional-dependencies"]
+
+    assert list(optional["nemo-evaluator-plugin"]) == [
+        "nemo-evaluator-sdk",
+        "nemo-platform-sdk",
+        "nmp-common",
+        "pydantic>=2.10.6",
+    ]
+    assert list(optional["services"]) == ["rich>=14.1.0", "nemo-platform[core-service]"]
+
+
 def test_process_bundle_packages_inherits_requested_metadata(tmp_path: Path, monkeypatch) -> None:
     wrapper_path = tmp_path / "packages/nemo_platform"
     plugin_path = tmp_path / "plugins/nemo-switchyard"
diff --git a/uv.lock b/uv.lock
index 82c2fec643..e682fbb541 100644
--- a/uv.lock
+++ b/uv.lock
@@ -3813,9 +3813,9 @@ source = { editable = "plugins/nemo-evaluator" }
 dependencies = [
     { name = "cloudpickle", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nemo-evaluator-sdk", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
-    { name = "nemo-platform", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nemo-platform-plugin", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
-    { name = "nmp-evaluator", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+    { name = "nemo-platform-sdk", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+    { name = "nmp-common", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "pydantic", extra = ["email"], marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "typer", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
 ]
@@ -3831,9 +3831,9 @@ dev = [
 requires-dist = [
     { name = "cloudpickle", specifier = ">=3.1.1" },
     { name = "nemo-evaluator-sdk", editable = "packages/nemo_evaluator_sdk" },
-    { name = "nemo-platform", editable = "packages/nemo_platform" },
     { name = "nemo-platform-plugin", editable = "packages/nemo_platform_plugin" },
-    { name = "nmp-evaluator", editable = "services/evaluator" },
+    { name = "nemo-platform-sdk", editable = "sdk/python/nemo-platform" },
+    { name = "nmp-common", editable = "packages/nmp_common" },
     { name = "pydantic", specifier = ">=2.10.6" },
     { name = "typer", specifier = ">=0.20.0" },
 ]
@@ -4003,7 +4003,6 @@ all = [
     { name = "nemoguardrails", extra = ["tracing"], marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "ngcsdk", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nmp-common", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
-    { name = "nmp-evaluator", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nvidia-nat-config-optimizer", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nvidia-nat-core", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nvidia-nat-langchain", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
@@ -4304,7 +4303,8 @@ nemo-evaluator-plugin = [
     { name = "cloudpickle", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nemo-evaluator-sdk", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nemo-platform-plugin", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
-    { name = "nmp-evaluator", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+    { name = "nemo-platform-sdk", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+    { name = "nmp-common", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "pydantic", extra = ["email"], marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "typer", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
 ]
@@ -4410,10 +4410,10 @@ plugins = [
     { name = "nemo-anonymizer", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nemo-evaluator-sdk", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nemo-platform-plugin", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
+    { name = "nemo-platform-sdk", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nemo-safe-synthesizer", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nemoguardrails", extra = ["tracing"], marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nmp-common", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
-    { name = "nmp-evaluator", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nvidia-nat-config-optimizer", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nvidia-nat-core", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nvidia-nat-langchain", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
@@ -4477,7 +4477,6 @@ services = [
     { name = "nemoguardrails", extra = ["tracing"], marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "ngcsdk", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nmp-common", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
-    { name = "nmp-evaluator", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nvidia-nat-config-optimizer", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nvidia-nat-core", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
     { name = "nvidia-nat-langchain", marker = "(platform_machine == 'arm64' and sys_platform == 'darwin') or (platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
@@ -4790,8 +4789,10 @@ requires-dist = [
     { name = "nemo-platform-sdk", marker = "extra == 'all'", editable = "sdk/python/nemo-platform" },
     { name = "nemo-platform-sdk", marker = "extra == 'core-service'", editable = "sdk/python/nemo-platform" },
     { name = "nemo-platform-sdk", marker = "extra == 'entities-service'", editable = "sdk/python/nemo-platform" },
+    { name = "nemo-platform-sdk", marker = "extra == 'nemo-evaluator-plugin'", editable = "sdk/python/nemo-platform" },
     { name = "nemo-platform-sdk", marker = "extra == 'nemo-platform-plugin'", editable = "sdk/python/nemo-platform" },
     { name = "nemo-platform-sdk", marker = "extra == 'nmp-common'", editable = "sdk/python/nemo-platform" },
+    { name = "nemo-platform-sdk", marker = "extra == 'plugins'", editable = "sdk/python/nemo-platform" },
     { name = "nemo-platform-sdk", marker = "extra == 'services'", editable = "sdk/python/nemo-platform" },
     { name = "nemo-safe-synthesizer", marker = "extra == 'all'", specifier = "==0.1.1" },
     { name = "nemo-safe-synthesizer", marker = "extra == 'nemo-safe-synthesizer-plugin'", specifier = "==0.1.1" },
@@ -4826,15 +4827,12 @@ requires-dist = [
     { name = "nmp-common", marker = "extra == 'nemo-agents-plugin'", editable = "packages/nmp_common" },
     { name = "nmp-common", marker = "extra == 'nemo-anonymizer-plugin'", editable = "packages/nmp_common" },
     { name = "nmp-common", marker = "extra == 'nemo-data-designer-plugin'", editable = "packages/nmp_common" },
+    { name = "nmp-common", marker = "extra == 'nemo-evaluator-plugin'", editable = "packages/nmp_common" },
     { name = "nmp-common", marker = "extra == 'nemo-safe-synthesizer-plugin'", editable = "packages/nmp_common" },
     { name = "nmp-common", marker = "extra == 'plugins'", editable = "packages/nmp_common" },
     { name = "nmp-common", marker = "extra == 'secrets-service'", editable = "packages/nmp_common" },
     { name = "nmp-common", marker = "extra == 'services'", editable = "packages/nmp_common" },
     { name = "nmp-common", marker = "extra == 'studio-service'", editable = "packages/nmp_common" },
-    { name = "nmp-evaluator", marker = "extra == 'all'", editable = "services/evaluator" },
-    { name = "nmp-evaluator", marker = "extra == 'nemo-evaluator-plugin'", editable = "services/evaluator" },
-    { name = "nmp-evaluator", marker = "extra == 'plugins'", editable = "services/evaluator" },
-    { name = "nmp-evaluator", marker = "extra == 'services'", editable = "services/evaluator" },
     { name = "nvidia-ml-py", marker = "extra == 'nemo-platform-sdk'", specifier = ">=13.0.0" },
     { name = "nvidia-ml-py", marker = "extra == 'nmp-common'", specifier = ">=13.0.0" },
     { name = "nvidia-nat-config-optimizer", marker = "extra == 'all'", specifier = ">=1.7.0,<1.8" },