Reproduction of Self-Distillation Fine-Tuning as a Reef weight-training recipe. SDFT learns from demonstrations on policy: the served model reads a demonstration in its prompt and is the teacher, the same model without it is the student, and the loss is the per-token KL between their next-token distributions along the student's own sample. The package holds the method; its report contract, a rollout's receipt plus the teacher's context (reef.core.reports.TeacherContextReport), is shared with SDPO.

• Paper: arXiv:2601.19897
• Reference implementation: idanshen/Self-Distillation at d77573212fa0; the recipe's processor maps onto its main.py (the demonstration prompt) and the Slime backend's distillation base (reef/train/slime_backend/distill/, which the sdft family configures) onto distil_trainer.py (the loss). Forward KL is the default, as the authors' 2026-04-07 note says the paper's results used it; reverse KL is a switch.
• Pins: slime pinned to THUDM/slime@41014d1f29e201137fdffce737bb8bac65bc5219 (via pyproject.toml dependency-groups.runtime)
• Claim scope: SDFT on a skill stream, the paper's Figure 3 protocol (Tool Use, then Science Q&A) against an SFT control on the same demonstrations. The CEO-Bench comparison is the roadmap's next item (#502).

sdft/
  recipe.py          SDFTRecipe: training spec, loss family "sdft"; its report contract is reef.core.reports.TeacherContextReport
  processor.py       the shared DistillProcessor with the demonstration appended to the recorded request
  objective.py       selects the sdft loss; the recipe binds the per-sample step schedule
  slime/             the loss family: SDFT's defaults and hook names on the Slime backend's distillation base
  examples/
    skill_stream/    Tool Use, then Science Q&A, through reef-eval: the paper's Figure 3 against an SFT control

The sdft recipe page covers the report contract, configuration and the driver flags, and Loss families describes how the family plugs into the Slime backend.