marin: require pydantic BaseModel for typed Artifact.from_path

Drops the dataclass branch from typed loading — `artifact_type` must now
be a pydantic BaseModel subclass, otherwise `from_path` raises TypeError.
All production call sites already pass pydantic types
(`NormalizedData`, `MinHashAttrData`, `FuzzyDupsAttrData`); test types
`TokenizeMetadata` / `TrainMetadata` are converted to BaseModel. Save
still accepts dataclasses since they round-trip fine through untyped
`from_path`.

Author: ravwojdyla

lib/marin/src/marin/execution/artifact.py

@@ -48,11 +48,9 @@ def from_path(
             with open_url(f"{base_path}/{cls.__artifact_file_name}", "rb") as fd:
                 if artifact_type is None:
                     return json.load(fd)
-                if issubclass(artifact_type, BaseModel):
-                    return artifact_type.model_validate_json(fd.read())
-                if is_dataclass(artifact_type):
-                    return artifact_type(**json.load(fd))  # type: ignore[not-callable]
-                raise ValueError(f"Unsupported artifact type: {artifact_type!r}")
+                if not issubclass(artifact_type, BaseModel):
+                    raise TypeError(f"artifact_type must be a pydantic BaseModel subclass, got {artifact_type!r}")
+                return artifact_type.model_validate_json(fd.read())
         except FileNotFoundError:
             return cls._from_executor_status(base_path, artifact_type)
 

tests/execution/test_step_runner.py

@@ -17,21 +17,20 @@
 from marin.execution.remote import RemoteCallable, remote
 from marin.execution.step_runner import StepRunner
 from marin.execution.step_spec import StepSpec
+from pydantic import BaseModel
 from rigging.filesystem import MARIN_CROSS_REGION_OVERRIDE_ENV
 
 # ---------------------------------------------------------------------------
 # Artifact types
 # ---------------------------------------------------------------------------
 
 
-@dataclass
-class TokenizeMetadata:
+class TokenizeMetadata(BaseModel):
     path: str
     num_tokens: int
 
 
-@dataclass
-class TrainMetadata:
+class TrainMetadata(BaseModel):
     tokens_seen: int
     checkpoint_path: str
 
@@ -158,18 +157,6 @@ def test_artifact_from_executor_status_non_success_raises(tmp_path: Path):
         Artifact.from_path(tmp_path.as_posix())
 
 
-def test_artifact_from_executor_status_relative_path(tmp_path: Path, monkeypatch):
-    """Fallback also works through MARIN_PREFIX-resolved relative paths."""
-    monkeypatch.setenv("MARIN_PREFIX", tmp_path.as_posix())
-    step_dir = tmp_path / "step_out"
-    step_dir.mkdir()
-    (step_dir / ".executor_status").write_text("SUCCESS")
-
-    loaded = Artifact.from_path("step_out")
-    assert isinstance(loaded, PathMetadata)
-    assert loaded.path == step_dir.as_posix()
-
-
 def test_artifact_save_and_load_untyped(tmp_path: Path):
     artifact = TokenizeMetadata(path="/tokenized", num_tokens=42)
     Artifact.save(artifact, tmp_path.as_posix())