Skip to main content
Start with SFT, GRPO, or evaluation. Use More algorithms when you need a variant or distillation objective beyond those foundations. Every loop follows this safety envelope:
On error or interruption:
Implement algorithms as caller-owned loops over Whitney primitives. There is no managed recipe submit endpoint and no Whitney SDK requirement for v1.

​
Runnable reference implementations

Every loop on this page and on More algorithms has a real, runnable counterpart in examples/training/http/ — not pseudocode. sl_loop.py and rl_loop.py are the two foundational loops; most algorithm variants are a single build_loss(sample, cycle, request, ctx) callable plugged into rl_loop.py’s --build-loss flag, ported from tinker-cookbook==0.5.5 (the same version pinned in Whitney’s own backend) where an upstream implementation exists, and cited to it in the code. loss_config values (PPO/GSPO/CISPO’s clip range, DPPO’s divergence thresholds) reaching the provider at all needs a recent fix in Whitney’s Modal trainer bridge — see the note at the top of More algorithms. Every stub file is still a real, non-allocating capability check (python <file>.py --provider ... --model ...) with a docstring explaining exactly what is missing — never a fabricated implementation.