diff --git a/CHANGELOG.md b/CHANGELOG.md
index 805411a..eedd50b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -16,7 +16,7 @@ All notable changes to this project will be documented in this file.
 - `MetricWriter` is IO-first; open a path with `MetricWriter.open()` rather than passing it to the constructor (#42)
 - `Metric.read` is now a thin wrapper over `MetricReader.open` and reads eagerly, returning a `list` instead of a lazy generator; it accepts `str` paths in addition to `Path` and gains transparent decompression and the `encoding` kwarg. Use `MetricReader.open` to stream metrics without holding them all in memory (#62)
 - `MetricReader.open` and `MetricWriter.open` eagerly validate that the path is readable/writable on context entry (#66)
-- Rename the `DelimitedList` converter to `DelimitedCollection` (#76)
+- Generalize the `DelimitedList` converter to `DelimitedCollection`, which now also handles `set`, `frozenset`, and `tuple` (including heterogeneous tuples such as `tuple[int, str]`) in addition to `list`. Set and frozenset output is sorted by serialized form for stable roundtrips (#76)
 
 ## [0.3.0] - 2026-05-12
 
diff --git a/fgmetric/_typing_extensions.py b/fgmetric/_typing_extensions.py
index cea3ce6..62d7134 100644
--- a/fgmetric/_typing_extensions.py
+++ b/fgmetric/_typing_extensions.py
@@ -178,6 +178,33 @@ def is_list(annotation: TypeAnnotation | None) -> bool:
     return has_origin(annotation, list)
 
 
+def is_collection(annotation: TypeAnnotation | None) -> bool:
+    """
+    Check if a type annotation is a delimited-collection type.
+
+    Matches parameterized `list`, `set`, `frozenset`, and `tuple` (including the optional and
+    higher-arity tuple forms). Mappings (`dict`, `Counter`) and bare, unparameterized collections
+    are not collections for this purpose.
+
+    Examples:
+        >>> is_collection(list[int])
+        True
+        >>> is_collection(set[str])
+        True
+        >>> is_collection(frozenset[int])
+        True
+        >>> is_collection(tuple[int, str])
+        True
+        >>> is_collection(tuple[int, ...] | None)
+        True
+        >>> is_collection(dict[str, int])
+        False
+        >>> is_collection(list)  # bare list, no type parameter
+        False
+    """
+    return any(has_origin(annotation, origin) for origin in (list, set, frozenset, tuple))
+
+
 def is_counter(annotation: TypeAnnotation | None) -> bool:
     """
     True if the type annotation is a Counter.
diff --git a/fgmetric/converters/_delimited_collection.py b/fgmetric/converters/_delimited_collection.py
index eed40fb..98047de 100644
--- a/fgmetric/converters/_delimited_collection.py
+++ b/fgmetric/converters/_delimited_collection.py
@@ -1,7 +1,8 @@
 from typing import Any
 from typing import ClassVar
-from typing import TypeAlias
 from typing import final
+from typing import get_args
+from typing import get_origin
 
 from pydantic import BaseModel
 from pydantic import FieldSerializationInfo
@@ -10,35 +11,43 @@
 from pydantic import field_serializer
 from pydantic import field_validator
 
-from fgmetric._typing_extensions import has_optional_elements
-from fgmetric._typing_extensions import is_list
-
-Fieldname: TypeAlias = str
-"""A pydantic model field's name."""
+from fgmetric._typing_extensions import is_collection
+from fgmetric._typing_extensions import is_optional
+from fgmetric._typing_extensions import unpack_optional
 
 
 # NB: Inheriting from BaseModel is necessary to declare field/model validators on the mixin, and
 # for the class-level validations defined in `__pydantic_init_subclass__` to work.
 class DelimitedCollection(BaseModel):
     """
-    Serialize and deserialize delimited lists of (de)serializable types.
+    Serialize and deserialize delimited collections of (de)serializable types.
 
-    When this mixin is added to `Metric`, fields annotated as `list[T]` will be read and written as
-    comma-delimited strings. During validation, a comma-delimited string will be split into a list
-    and its elements validated as instances of `T`. During serialization, the list elements will be
-    serialized to string and then joined into a comma-delimited string.
+    When this mixin is added to `Metric`, fields annotated as `list[T]`, `set[T]`, `frozenset[T]`,
+    or `tuple[...]` will be read and written as delimited strings. During validation, a delimited
+    string is split into its elements, which are validated as instances of the declared element
+    type(s); Pydantic performs the container coercion (list, set, frozenset, or tuple). During
+    serialization, the elements are serialized to string and joined back into a delimited string.
 
-    The list type `T` may be any serializable type. The field may be annotated as `list[T]` or
-    `list[T] | None` - as with any primitive type, `None` will be validated from and serialized to
-    the empty string.
+    The element type(s) may be any serializable type. A field may be annotated as `list[T]`,
+    `set[T]`, `frozenset[T]`, the variadic `tuple[T, ...]`, or the fixed-arity, heterogeneous
+    `tuple[T1, T2, ...]` (whose arity and per-position types Pydantic validates). Any of these may
+    also be made optional (e.g. `list[T] | None`); as with any primitive type, `None` is validated
+    from and serialized to the empty string.
 
     The delimiter may be configured by specifying the `collection_delimiter` class variable when
     declaring a model.
 
     Note:
-        Roundtrips are lossy if list elements contain the delimiter character. For example, with the
+        `list` and `tuple` are ordered, so their element order is preserved. `set` and `frozenset`
+        are unordered (and string hashing is per-process salted), so their elements are serialized
+        sorted by their serialized form, which keeps the output stable across runs and roundtrips.
+
+    Note:
+        Roundtrips are lossy if elements contain the delimiter character. For example, with the
         default comma delimiter, `["a,b", "c"]` serializes to `"a,b,c"` and deserializes back to
-        `["a", "b", "c"]`. Avoid using delimiters that may appear in element values.
+        `["a", "b", "c"]`. Avoid using delimiters that may appear in element values. Collections
+        are also flat: nested collections (e.g. `list[list[int]]`) are not supported, because a
+        flat delimited string cannot be unambiguously re-nested.
 
     Examples:
         Basic usage — comma delimiter (default):
@@ -51,6 +60,25 @@ class MyMetric(Metric):
         MyMetric(tags=[1, 2, 3]).model_dump()  # -> {"tags": "1,2,3"}
         ```
 
+        Sets and frozensets — output is sorted by serialized form:
+
+        ```python
+        class MyMetric(Metric):
+            tags: set[int]  # "3,1,2" becomes {1, 2, 3}
+
+        MyMetric.model_validate({"tags": "3,1,2"}).tags        # -> {1, 2, 3}
+        MyMetric(tags={3, 1, 2}).model_dump()  # -> {"tags": "1,2,3"}
+        ```
+
+        Heterogeneous tuples — arity and per-position types are validated:
+
+        ```python
+        class MyMetric(Metric):
+            point: tuple[int, str]  # "1,foo" becomes (1, "foo")
+
+        MyMetric.model_validate({"point": "1,foo"}).point     # -> (1, "foo")
+        ```
+
         Custom delimiter:
 
         ```python
@@ -62,7 +90,7 @@ class MyMetric(Metric):
         MyMetric(tags=[1, 2, 3]).model_dump()  # -> {"tags": "1;2;3"}
         ```
 
-        Optional list field — the whole field may be absent:
+        Optional collection field — the whole field may be absent:
 
         ```python
         class MyMetric(Metric):
@@ -72,7 +100,7 @@ class MyMetric(Metric):
         MyMetric(tags=None).model_dump()   # -> {"tags": None}
         ```
 
-        List with optional elements — individual elements may be absent:
+        Collection with optional elements — individual elements may be absent:
 
         ```python
         class MyMetric(Metric):
@@ -84,29 +112,23 @@ class MyMetric(Metric):
     """
 
     collection_delimiter: ClassVar[str] = ","
-    _list_fieldnames: ClassVar[set[str]]
-    _optional_element_fieldnames: ClassVar[set[str]]
+    _collection_fieldnames: ClassVar[set[str]]
+    _uniform_optional_element_fieldnames: ClassVar[set[str]]
+    _tuple_optional_positions: ClassVar[dict[str, tuple[bool, ...]]]
 
     @classmethod
     def __pydantic_init_subclass__(cls, **kwargs: Any) -> None:
         """
-        Validations of the user-defined model.
+        Validate the collection delimiter and classify the model's collection fields.
 
         1. The collection delimiter must be a single character.
-        2. The names of all fields annotated as `list[T]` or `list[T] | None` are stored in the
-           private `_list_fieldnames` class variable.
+        2. Each `list`/`set`/`frozenset`/`tuple` field is recorded, along with the
+           per-element/per-position optionality needed to map empty cells to `None`.
         """
         super().__pydantic_init_subclass__(**kwargs)
 
         cls._require_single_character_collection_delimiter()
-        cls._list_fieldnames = {
-            name for name, info in cls.model_fields.items() if is_list(info.annotation)
-        }
-        cls._optional_element_fieldnames = {
-            name
-            for name, info in cls.model_fields.items()
-            if has_optional_elements(info.annotation)
-        }
+        cls._classify_collection_fields()
 
     @classmethod
     def _require_single_character_collection_delimiter(cls) -> None:
@@ -117,51 +139,110 @@ def _require_single_character_collection_delimiter(cls) -> None:
                 f" got: {cls.collection_delimiter!r}"
             )
 
+    @classmethod
+    def _classify_collection_fields(cls) -> None:
+        """Scan the model's fields once, recording each collection field's behavior."""
+        cls._collection_fieldnames = set()
+        cls._uniform_optional_element_fieldnames = set()
+        cls._tuple_optional_positions = {}
+
+        for name, info in cls.model_fields.items():
+            annotation = info.annotation
+            if not is_collection(annotation):
+                continue
+
+            cls._collection_fieldnames.add(name)
+
+            inner = _strip_optional(annotation)
+            args = get_args(inner)
+
+            # A fixed-arity tuple (e.g. `tuple[int, str]`) has per-position element types; every
+            # other collection — including the variadic `tuple[T, ...]` — has a single, uniform
+            # element type.
+            is_fixed_tuple = (
+                get_origin(inner) is tuple and args != () and not _is_variadic_tuple_args(args)
+            )
+            if is_fixed_tuple:
+                cls._tuple_optional_positions[name] = tuple(is_optional(arg) for arg in args)
+            elif args and is_optional(args[0]):
+                cls._uniform_optional_element_fieldnames.add(name)
+
     @final
     @field_validator("*", mode="before")
     @classmethod
-    def _split_lists(cls, value: Any, info: ValidationInfo) -> Any:
-        """Split any fields annotated as `list[T]` on a comma delimiter."""
-        if isinstance(value, str) and cls._is_list_field(info.field_name):
-            if value:
-                value = value.split(cls.collection_delimiter)
+    def _split_collections(cls, value: Any, info: ValidationInfo) -> Any:
+        """Split any collection field into the intermediate list Pydantic then coerces."""
+        if isinstance(value, str) and cls._is_collection_field(info.field_name):
+            return cls._split_collection(value, info.field_name)
+        return value
+
+    @final
+    @classmethod
+    def _split_collection(cls, value: str, name: str | None) -> list[Any]:
+        """Split a collection cell into a flat list of (string-or-`None`) elements."""
+        if not value:
+            return []
 
-                # Convert empty strings to None for list[T | None] fields
-                if info.field_name in cls._optional_element_fieldnames:
-                    value = [None if el == "" else el for el in value]
+        elements: list[Any] = value.split(cls.collection_delimiter)
 
-            else:
-                value = []
+        # Map empty elements to `None` where the corresponding element type is optional.
+        if name in cls._uniform_optional_element_fieldnames:
+            elements = [None if element == "" else element for element in elements]
+        elif name in cls._tuple_optional_positions:
+            mask = cls._tuple_optional_positions[name]
+            elements = [
+                None if (index < len(mask) and mask[index] and element == "") else element
+                for index, element in enumerate(elements)
+            ]
 
-        return value
+        return elements
 
     @final
     @field_serializer("*", mode="wrap")
-    def _join_lists(
+    def _join_collections(
         self,
         value: Any,
-        nxt: SerializerFunctionWrapHandler,  # noqa: ARG002
+        nxt: SerializerFunctionWrapHandler,
         info: FieldSerializationInfo,
     ) -> Any:
-        """Join any fields annotated as `list[T]` with a delimiter."""
-        if isinstance(value, list) and self._is_list_field(info.field_name):
-            # Let the default serializer handle each item first. This should return a list of
-            # serialized values, applying default serialization to each list element.
-            serialized_value = nxt(value)
-
-            if isinstance(serialized_value, list):
-                # If the handler returned a list, join it. (This is the expected branch.)
-                # Also serialize `None` back to empty string.
-                elements = ["" if item is None else str(item) for item in serialized_value]
-                return self.collection_delimiter.join(elements)
-            else:
-                # If the handler already serialized to something else (unlikely), return as-is.
-                return serialized_value
-
+        """Join any collection field back into a single delimited string."""
+        if self._is_collection_field(info.field_name) and isinstance(
+            value, (list, set, frozenset, tuple)
+        ):
+            return self._join_collection(value, nxt)
         return nxt(value)
 
+    @final
+    def _join_collection(self, value: Any, nxt: SerializerFunctionWrapHandler) -> Any:
+        """Serialize each element, then join into a delimited string."""
+        # Let the default serializer handle each element first, applying any per-element custom
+        # serialization. This returns the elements as an iterable (a list in JSON mode).
+        serialized = nxt(value)
+        if not isinstance(serialized, (list, set, frozenset, tuple)):
+            # If the handler already produced something else (unlikely), return it as-is.
+            return serialized
+
+        elements = ["" if item is None else str(item) for item in serialized]
+
+        # `list`/`tuple` are ordered, so preserve order. `set`/`frozenset` are unordered (and
+        # string hashing is per-process salted), so sort by serialized form for a stable roundtrip.
+        if isinstance(value, (set, frozenset)):
+            elements.sort()
+
+        return self.collection_delimiter.join(elements)
+
     @final
     @classmethod
-    def _is_list_field(cls, field_name: str | None) -> bool:
-        """True if the field is annotated as `list[T]` on the class model."""
-        return field_name is not None and field_name in cls._list_fieldnames
+    def _is_collection_field(cls, field_name: str | None) -> bool:
+        """True if the field is annotated as a delimited collection on the class model."""
+        return field_name is not None and field_name in cls._collection_fieldnames
+
+
+def _strip_optional(annotation: Any) -> Any:
+    """Return the inner type of an optional annotation, or the annotation unchanged."""
+    return unpack_optional(annotation) if is_optional(annotation) else annotation
+
+
+def _is_variadic_tuple_args(args: tuple[Any, ...]) -> bool:
+    """True if a tuple's type arguments describe a variadic `tuple[T, ...]`."""
+    return len(args) == 2 and args[1] is Ellipsis
diff --git a/fgmetric/metric.py b/fgmetric/metric.py
index d35559f..edc4aae 100644
--- a/fgmetric/metric.py
+++ b/fgmetric/metric.py
@@ -35,12 +35,12 @@ class Metric(
     1. **Null sentinels.** Empty strings in Optional fields are converted to `None` before field
        validation. The sentinel values converted to `None` can be overridden by the
        `null_sentinels` class variable.
-    2. **Delimited lists.** Any field typed as `list[T]` will be parsed from and serialized to a
-       delimited string. The list delimiter may be controlled by the `collection_delimiter` class
-       variable.
+    2. **Delimited collections.** Any field typed as `list[T]`, `set[T]`, `frozenset[T]`, or
+       `tuple[...]` will be parsed from and serialized to a delimited string. The delimiter may be
+       controlled by the `collection_delimiter` class variable.
 
     Class Variables:
-        collection_delimiter: A single-character delimiter used to split and join `list` fields
+        collection_delimiter: A single-character delimiter used to split and join collection fields
             during serialization/deserialization.
         null_sentinels: The set of input strings that should be treated as `None` on Optional
             fields during validation. Defaults to `frozenset({""})`.
diff --git a/tests/test_delimited_collection.py b/tests/test_delimited_collection.py
index 55820d5..861fe6c 100644
--- a/tests/test_delimited_collection.py
+++ b/tests/test_delimited_collection.py
@@ -7,6 +7,7 @@
 from fgmetric import Metric
 from fgmetric import MetricReader
 from fgmetric import MetricWriter
+from fgmetric.converters import DelimitedCollection
 
 
 def test_comma_delimited_list(tmp_path: Path) -> None:
@@ -16,8 +17,8 @@ class FakeMetric(Metric):
         name: str
         values: list[int]
 
-    assert FakeMetric._list_fieldnames == {"values"}
-    assert FakeMetric._is_list_field("values")
+    assert FakeMetric._collection_fieldnames == {"values"}
+    assert FakeMetric._is_collection_field("values")
 
     # Test reading
     fpath_to_read = tmp_path / "test.txt"
@@ -111,8 +112,8 @@ class FakeMetric(Metric):
         name: str
         values: list[int] | None
 
-    assert FakeMetric._list_fieldnames == {"values"}
-    assert FakeMetric._is_list_field("values")
+    assert FakeMetric._collection_fieldnames == {"values"}
+    assert FakeMetric._is_collection_field("values")
 
     # Test reading
     fpath_to_read = tmp_path / "test.txt"
@@ -184,3 +185,125 @@ class FakeMetric(Metric):
 
     with MetricReader.open(FakeMetric, p) as reader:
         assert list(reader) == [metric]
+
+
+def test_metric_is_built_on_delimited_collection() -> None:
+    """`DelimitedCollection` is the canonical mixin; `Metric` is built on it."""
+
+    class FakeMetric(Metric):
+        values: list[int]
+
+    assert issubclass(Metric, DelimitedCollection)
+    assert issubclass(FakeMetric, DelimitedCollection)
+
+
+def test_set_field_roundtrip() -> None:
+    """A set[T] field parses from and serializes to a delimited string."""
+
+    class FakeMetric(Metric):
+        values: set[int]
+
+    assert FakeMetric._collection_fieldnames == {"values"}
+
+    m = FakeMetric.model_validate({"values": "3,1,2"})
+    assert m.values == {1, 2, 3}
+
+    # Sets have no order, so output is sorted by serialized form for a stable roundtrip.
+    assert m.model_dump()["values"] == "1,2,3"
+
+
+def test_set_output_is_sorted_stably() -> None:
+    """Set serialization is stable: sorted by serialized element form regardless of insertion."""
+
+    class FakeMetric(Metric):
+        values: set[str]
+
+    assert FakeMetric(values={"banana", "apple", "cherry"}).model_dump()["values"] == (
+        "apple,banana,cherry"
+    )
+
+
+def test_empty_set_field() -> None:
+    """An empty cell parses to an empty set and serializes back to an empty cell."""
+
+    class FakeMetric(Metric):
+        values: set[int]
+
+    m = FakeMetric.model_validate({"values": ""})
+    assert m.values == set()
+    assert m.model_dump()["values"] == ""
+
+
+def test_frozenset_field_roundtrip() -> None:
+    """A frozenset[T] field parses from and serializes to a sorted delimited string."""
+
+    class FakeMetric(Metric):
+        values: frozenset[int]
+
+    assert FakeMetric._collection_fieldnames == {"values"}
+
+    m = FakeMetric.model_validate({"values": "3,1,2"})
+    assert m.values == frozenset({1, 2, 3})
+    assert m.model_dump()["values"] == "1,2,3"
+
+
+def test_homogeneous_tuple_field_roundtrip() -> None:
+    """A variadic tuple[T, ...] field parses and serializes preserving order."""
+
+    class FakeMetric(Metric):
+        values: tuple[int, ...]
+
+    assert FakeMetric._collection_fieldnames == {"values"}
+
+    m = FakeMetric.model_validate({"values": "3,1,2"})
+    assert m.values == (3, 1, 2)
+    # Tuples are ordered, so order is preserved (not sorted).
+    assert m.model_dump()["values"] == "3,1,2"
+
+
+def test_heterogeneous_tuple_field_roundtrip() -> None:
+    """A fixed-arity tuple[int, str] field validates arity and per-position types."""
+
+    class FakeMetric(Metric):
+        values: tuple[int, str]
+
+    m = FakeMetric.model_validate({"values": "1,foo"})
+    assert m.values == (1, "foo")
+    assert m.model_dump()["values"] == "1,foo"
+
+
+def test_heterogeneous_tuple_arity_is_validated() -> None:
+    """A fixed-arity tuple rejects the wrong number of elements."""
+
+    class FakeMetric(Metric):
+        values: tuple[int, str]
+
+    with pytest.raises(ValueError):
+        FakeMetric.model_validate({"values": "1,foo,extra"})
+
+
+def test_set_with_optional_elements() -> None:
+    """A set[T | None] field maps empty elements to None."""
+
+    class FakeMetric(Metric):
+        values: set[int | None]
+
+    m = FakeMetric.model_validate({"values": "1,,3"})
+    assert m.values == {1, None, 3}
+    # Sorted by serialized form: "" (None) sorts before "1" and "3".
+    assert m.model_dump()["values"] == ",1,3"
+
+
+def test_tuple_with_optional_position() -> None:
+    """A fixed tuple maps an empty element to None only at optional positions."""
+
+    class FakeMetric(Metric):
+        values: tuple[str, int | None]
+
+    # Position 0 (str) keeps the empty string; position 1 (int | None) becomes None.
+    m = FakeMetric.model_validate({"values": ",5"})
+    assert m.values == ("", 5)
+
+    m2 = FakeMetric.model_validate({"values": "name,"})
+    assert m2.values == ("name", None)
+    assert m2.model_dump()["values"] == "name,"
diff --git a/tests/test_typing_extensions.py b/tests/test_typing_extensions.py
index 6cdbe3a..b99bb6d 100644
--- a/tests/test_typing_extensions.py
+++ b/tests/test_typing_extensions.py
@@ -1,3 +1,4 @@
+from collections import Counter
 from typing import Optional
 from typing import Union
 
@@ -6,6 +7,7 @@
 from fgmetric._typing_extensions import TypeAnnotation
 from fgmetric._typing_extensions import has_optional_elements
 from fgmetric._typing_extensions import has_origin
+from fgmetric._typing_extensions import is_collection
 from fgmetric._typing_extensions import is_list
 from fgmetric._typing_extensions import is_optional
 from fgmetric._typing_extensions import unpack_optional
@@ -77,6 +79,38 @@ def test_is_not_list(annotation: TypeAnnotation) -> None:
     assert not is_list(annotation)
 
 
+@pytest.mark.parametrize(
+    "annotation",
+    [
+        list[int],
+        set[str],
+        frozenset[int],
+        tuple[int, ...],
+        tuple[int, str],
+        Optional[list[int]],
+        set[str] | None,
+    ],
+)
+def test_is_collection(annotation: TypeAnnotation) -> None:
+    """Should identify list/set/frozenset/tuple types, even within an Optional."""
+    assert is_collection(annotation)
+
+
+@pytest.mark.parametrize(
+    "annotation",
+    [
+        str,
+        str | None,
+        dict[str, int],
+        Counter[str],
+        list,
+    ],
+)
+def test_is_not_collection(annotation: TypeAnnotation) -> None:
+    """Should reject mappings, bare collections, and non-collection types."""
+    assert not is_collection(annotation)
+
+
 @pytest.mark.parametrize(
     "annotation,collection_type",
     [