Training runtimes

The catalog runtimes in project nodus; reference one as spec.runtime: nodus/<name> on a TrainingJob.

Runtime	Task	Modes	Summary
`distill`	Distill	Train	Knowledge distillation from a teacher model; lmbda 0 is offline distillation
`dpo`	DPO	Train	Direct preference optimization on prompt, chosen and rejected rows
`evaluate`	Evaluate	Evaluate	Benchmark evaluation on standard tasks, or over an Environment’s held-out tasks
`grpo-lora`	GRPO	Train, Evaluate	GRPO reinforcement learning with a LoRA adapter over an Environment, graded by the platform
`kto`	KTO	Train	Kahneman-Tversky optimization on unpaired desirable and undesirable completions
`orpo`	ORPO	Train	Odds-ratio preference optimization: SFT and preference alignment in one pass, no reference model
`reward-model`	Reward	Train	Reward-model training: a one-logit head scored on chosen versus rejected
`sft`	SFT	Train	Supervised fine-tuning (full, LoRA, QLoRA) and pretraining with task: Pretrain