sft

SFT - supervised fine-tuning on the SDK training loop.

All the composer plumbing (backend, step/save cadence, evaluators, loop wiring, artifacts) lives in :class:~evsys_sdk.algorithms.base.BaseAlgorithm. SFT only has to say how a batch is built:

• :meth:setup - standardize rows → typed :class:ChatMessagesRow, tokenize to a static list of :class:tinker.Datum with assistant-span loss masks.
• :meth:build_batch - slice batch_size datums per step (wrapping the dataset so num_steps can exceed one epoch).
• :meth:step_metrics - train_mean_nll from the per-position logprobs.

Researchers wanting a one-line tweak (focal loss, custom loss, extra metrics) subclass SFT and override build_batch / step_metrics - no SDK change needed.