diff --git a/tools/datasets/__init__.py b/tools/datasets/__init__.py
index faf5e8d96..64940ebc1 100644
--- a/tools/datasets/__init__.py
+++ b/tools/datasets/__init__.py
@@ -22,9 +22,9 @@
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Literal
 
+from tools import fs
 from tools.codemod import ruff
 from tools.datasets.npm import Npm
-from tools.fs import REPO_ROOT
 from tools.schemapi import utils
 
 if TYPE_CHECKING:
@@ -60,7 +60,7 @@ class Application:
     Parameters
     ----------
     out_dir_tools, out_dir_altair
-        Directories to store ``.parquet`` metadata files.
+        Directories to store metadata files.
     out_fp_typing
         Path to write metadata-derived typing module.
 
@@ -72,7 +72,7 @@ class Application:
     def __init__(
         self, out_dir_tools: Path, out_dir_altair: Path, out_fp_typing: Path
     ) -> None:
-        out_dir_tools.mkdir(exist_ok=True)
+        fs.mkdir(out_dir_tools)
         METADATA = "metadata"
         self.paths = types.MappingProxyType["_PathAlias", Path](
             {
@@ -102,7 +102,7 @@ def refresh(
         include_typing
             Regenerate ``altair.datasets._typing``.
         frozen
-            Don't perform any requests or attempt to check for new versions.
+            Don't perform any requests.
 
             .. note::
                 **Temporary** measure to work from ``main`` until `vega-datasets@3`_.
@@ -123,20 +123,28 @@ def refresh(
 
     def reset(self) -> None:
         """Remove all metadata files."""
-        for fp in self.paths.values():
-            fp.unlink(missing_ok=True)
+        fs.rm(*self.paths.values())
 
     def read(self, name: _PathAlias, /) -> pl.DataFrame:
         """Read existing metadata from file."""
-        import polars as pl
-
-        return pl.read_parquet(self.paths[name])
+        return self.scan(name).collect()
 
     def scan(self, name: _PathAlias, /) -> pl.LazyFrame:
         """Scan existing metadata from file."""
         import polars as pl
 
-        return pl.scan_parquet(self.paths[name])
+        fp = self.paths[name]
+        if fp.suffix == ".parquet":
+            return pl.scan_parquet(fp)
+        elif ".csv" in fp.suffixes:
+            return pl.scan_csv(fp)
+        elif ".json" in fp.suffixes:
+            return pl.read_json(fp).lazy()
+        else:
+            msg = (
+                f"Unable to read {fp.name!r} as tabular data.\nSuffixes: {fp.suffixes}"
+            )
+            raise NotImplementedError(msg)
 
     def write_csv_gzip(self, frame: pl.DataFrame | pl.LazyFrame, fp: Path, /) -> None:
         """
@@ -152,8 +160,7 @@ def write_csv_gzip(self, frame: pl.DataFrame | pl.LazyFrame, fp: Path, /) -> Non
         """
         if fp.suffix != ".gz":
             fp = fp.with_suffix(".csv.gz")
-        if not fp.exists():
-            fp.touch()
+        fp.touch()
         df = frame.lazy().collect()
         buf = BytesIO()
         with gzip.GzipFile(fp, mode="wb", mtime=0) as f:
@@ -169,16 +176,13 @@ def write_json_gzip(self, obj: Any, fp: Path, /) -> None:
         """
         if fp.suffix != ".gz":
             fp = fp.with_suffix(".json.gz")
-        if not fp.exists():
-            fp.touch()
-
+        fp.touch()
         with gzip.GzipFile(fp, mode="wb", mtime=0) as f:
             f.write(json.dumps(obj).encode())
 
     def write_parquet(self, frame: pl.DataFrame | pl.LazyFrame, fp: Path, /) -> None:
         """Write ``frame`` to ``fp``, with some extra safety."""
-        if not fp.exists():
-            fp.touch()
+        fp.touch()
         df = frame.lazy().collect()
         df.write_parquet(fp, compression="zstd", compression_level=17)
 
@@ -233,7 +237,7 @@ def generate_typing(self, dpkg: datapackage.DataPackage) -> None:
         ruff.write_lint_format(self.paths["typing"], contents)
 
 
-_alt_datasets = REPO_ROOT / "altair" / "datasets"
+_alt_datasets = fs.REPO_ROOT / "altair" / "datasets"
 app = Application(
     Path(__file__).parent / "_metadata",
     _alt_datasets / "_metadata",