nodus.recipes.finetune

Fine-tuning, preference training, distillation and pretraining recipes on the catalog runtimes.

Each function returns a TrainingJob builder; nothing runs until .preview() or .run(gpu=…, max_cost=…). Keyword arguments not named here are runtime parameters in snake_case (learning_rate=1e-5 is parameters.learningRate), validated by the server against the runtime’s schema.

sft = finetune.sft(model="Qwen/Qwen3-1.7B", data=nodus.Volume.from_name("support-chats"),
                   lora=finetune.LoRA(r=16), max_steps=2000).run(gpu="H100", max_cost=20)
pt = finetune.pretrain(model="Qwen/Qwen3-0.6B", data=corpus, initialization="Scratch")
big = finetune.sft(model=m, data=d, distributed="ZeRO3").run(gpu="H100:8", nodes=2, max_cost=200)

Exports

• Data: import with from nodus.recipes.finetune import Data
• LoRA: import with from nodus.recipes.finetune import LoRA

distill

distill(*, student: Any, teacher: Any, data: Any, temperature: float | None = None, alpha: float | None = None, lmbda: float | None = None, **params: Any) -> Any

Knowledge distillation from a teacher model (distill); lmbda=0 is offline KD on the dataset’s text.

dpo

dpo(*, model: Any, data: Any, beta: float | None = None, **params: Any) -> Any

Direct preference optimisation on prompt, chosen and rejected rows (dpo).

kto

kto(*, model: Any, data: Any, beta: float | None = None, **params: Any) -> Any

KTO on prompt, completion and a boolean label (kto).

orpo

orpo(*, model: Any, data: Any, beta: float | None = None, **params: Any) -> Any

Odds-ratio preference optimisation, no reference model (orpo).

pretrain

pretrain(*, model: Any, data: Any, initialization: str = 'Continued', packing: bool = True, append_eos: bool | None = None, **params: Any) -> Any

Causal-LM pretraining on raw text (sft, task: Pretrain).

initialization="Continued" keeps the model’s weights; "Scratch" uses only its config and tokenizer (optionally resized by architecture={...}) and starts from random weights. append_eos=False preserves corpus boundaries instead of appending EOS to each text row. Omit it to retain the runtime’s default. Validation and packing still use the runtime’s parameters.

reward

reward(*, model: Any, data: Any, **params: Any) -> Any

A reward model from chosen and rejected pairs (reward-model).

sft

sft(*, model: Any, data: Any, lora: LoRA | None = None, max_steps: int | None = None, **params: Any) -> Any

Supervised fine-tuning on prompt and completion (or chat messages) rows (sft).