EvSys

TrainingBatch

One step's worth of training data + loss spec.

Constructed by a :class:StepBuilder per step. The loop hands data to the backend along with loss_fn and loss_fn_config.

Attributes

attributedatalist[tinker.Datum]
attributeloss_fntinker.types.LossFnType | LossCallable

Either a tinker-recognised string name ("cross_entropy" / "importance_sampling" / ...) for a server-side loss, OR a Python callable for a client-side custom loss. The loop dispatches the right backend method based on the type.

attributeloss_fn_configdict[str, Any] | None
= None

Only used when loss_fn is a string.

attributemetricsdict[str, float]
= field(default_factory=dict)

Algorithm-precomputed per-step metrics (e.g. teacher entropy, reward stats). Merged into the per-step log row.

Functions

func__init__(self, data, loss_fn, loss_fn_config=None, metrics=dict()) -> None
paramself
paramdatalist[tinker.Datum]
paramloss_fntinker.types.LossFnType | LossCallable
paramloss_fn_configdict[str, Any] | None
= None
parammetricsdict[str, float]
= dict()

Returns

None

On this page