slicktune.types

Shared base abstractions for slick-tune.

Classes

Objective	Base training objective (what the model learns).
Strategy	Base parameter-update strategy (how weights change).

Module Contents

class slicktune.types.Objective

Bases: abc.ABC

Base training objective (what the model learns).

Parameters:

name (str) – Objective identifier (set by subclasses, e.g. "sft").

__slots__ = ()

name: str

abstract required_columns() → list[str]

Return dataset column names required by this objective.

Returns:

list[str] – Column names that must exist on the training dataset.

class slicktune.types.Strategy

Bases: abc.ABC

Base parameter-update strategy (how weights change).

Parameters:

name (str) – Strategy identifier (set by subclasses, e.g. "lora").

__slots__ = ()

abstract apply(model: Any) → Any

Wrap or configure model for this strategy.

Parameters:

model (Any) – Base causal-LM module (typically a Hugging Face model).

Returns:

Any – Model ready for training under this strategy.

abstract load_kwargs() → dict[str, Any]

Return kwargs for from_pretrained that this strategy requires.

Returns:

dict[str, Any] – Keyword arguments passed when loading the base model.

name: str