base

BaseAlgorithm - shared base for the SDK's training algorithms.

Every training algorithm (sft / sdft / rl) used to copy-paste the same composer body: allocate the tinker backend, resolve the step count + save cadence, log hyperparams, build the in-loop evaluators, wire a :class:~evsys_sdk.training.loop.TrainingLoop, run it, and record the checkpoint artifacts. The only thing that genuinely differed was how each step's batch is built. This base owns all of the shared plumbing; a concrete algorithm overrides just the per-algorithm pieces.

The base is itself a StepBuilder - it implements build_batch and step_metrics and hands self to the loop, so there is no separate StepBuilder object to construct. A subclass overrides:

• :meth:setup - one-time prep before the loop. Stash per-algorithm state on self (tokenized datums for SFT; dataset + teacher client + sampler provider for SDFT/RL) and set self._steps_per_epoch.
• :meth:build_batch - produce the :class:TrainingBatch for one step. The loss spec rides on the returned batch (loss_fn is not a separate override). SFT slices its static datums; RL/SDFT roll out on-policy here.
• :meth:step_metrics - (optional) per-step metrics from the forward-backward result. Defaults to \{\}.

Researchers wanting a one-line tweak (focal loss, an extra metric, a custom loss) subclass the concrete algorithm and override build_batch / step_metrics - no SDK change needed.