sdft

SDFT - self-distillation fine-tuning on the SDK training loop.

All the composer plumbing lives in :class:~evsys_sdk.algorithms.base.BaseAlgorithm; SDFT supplies the two per-algorithm pieces:

• :meth:setup - parse rows → :class:PromptExample dataset, build the frozen teacher sampling client, and stash a per-step student sampler provider (snapshots the current weights each step so rollouts stay on-policy - cookbook parity without monkey-patching sdft.train_step).
• :meth:build_batch - on-policy student rollout → teacher topK score → CE distillation Datums. :meth:step_metrics adds train/mean_loss from the forward-backward output.