GitHub - arvidma/dataclasses: Vibe coded backports of 3.8+ dataclass features to 3.6

Adds most dataclass features from Python versions after 3.7 to the excellent but long-since abandoned backport by Eric V. Smith (https://github.com/ericvsmith/dataclasses/).

For those of us still stuck writing Python for systems tied hard to the oldest LTS releases of Ubuntu and RedHat.

Matrix of backported and not-yet-backported features:

Python version	Feature	Status
3.8	`InitVar[T]` generic syntax	Backported
3.8	`replace()` handling of `InitVar` fields with defaults	Backported
3.10	`kw_only` parameter and `KW_ONLY` sentinel	Backported
3.10	`match_args` parameter (`__match_args__` generation)	Backported
3.10	`slots` parameter (`__slots__` generation)	Backported
3.11	`weakref_slot` parameter	Backported
3.11	Frozen+slots pickling (`__getstate__`/`__setstate__`)	Backported
3.11	Slots with `init=False` default fields (bpo-44649)	Backported
3.11–3.12	`fields()` traceback improvements	Backported
3.11–3.12	Mutable default value validation	Backported
3.12	Field named `BUILTINS` in frozen classes (gh-96151)	Backported
3.12	Special underscore field names like `_dflt_x` (gh-98886)	Backported
3.12	Inherited `__dict__` slot handling	Backported
3.12–3.13	Optimized `__eq__` and `__repr__` recursion guards	Backported
3.12	`asdict()`/`astuple()` atomic-type fast paths (gh-91896)	Backported
3.13	Single-`exec` method generation (`_FuncBuilder`, gh-107550)	Backported
3.13	Non-frozen subclass of frozen: mutable attr set/delete	Backported
3.14	`decorator` parameter for `make_dataclass()`	Backported
3.12	`module` parameter for `make_dataclass()`	Not backported
3.12	`defaultdict` support in `asdict()`/`astuple()`	Not backported
3.13	`__replace__()` method (`copy.replace()` support)	Not backported

Usage

Drop the single dataclasses.py file into your project (vendoring), or add it to your Python path. It is a drop-in replacement for the standard library module and can transparently replace the stdlib if running on a Python version that doesn't support some feature you like.

Supports Python 3.6 and later from a single codebase — the only version-sensitive code is the ClassVar detection, which adapts to the typing internals of the running interpreter. The test suite runs in CI on Python 3.6 through 3.13.

Note: don't mix this module with the stdlib dataclasses in the same class. InitVar, Field, field() and the KW_ONLY sentinel are compared by identity, so e.g. a stdlib InitVar annotation on a class decorated with this module's @dataclass will not be recognized. Import everything from one module, as in the example below.

try:
    from dataclasses import dataclass, field, KW_ONLY
except ImportError:
    from myproj.vendor.dataclasses import dataclass, field, KW_ONLY  # type: ignore[no-redef]

@dataclass(slots=True, kw_only=True)
class Point:
    x: float
    y: float
    z: float = 0.0

Full `@dataclass` decorator signature

@dataclass(
    init=True,           # 3.7
    repr=True,           # 3.7
    eq=True,             # 3.7
    order=False,         # 3.7
    unsafe_hash=False,   # 3.7
    frozen=False,        # 3.7
    match_args=True,     # 3.10
    kw_only=False,       # 3.10
    slots=False,         # 3.10
    weakref_slot=False,  # 3.11
)

Usage

Full @dataclass decorator signature

Full `@dataclass` decorator signature