Recipe
recipe.yaml is the durable path map for a run — method, data paths, eval gate. Edit it by hand or author with the SDK (Aquin.define / Run); both rewrite/read the same file. The kernel and CLI consume the map; they do not care which surface wrote it.
recipe.yaml · tabular
family: tabular with method linear | logistic | ridge | lasso | elasticnet | tree | forest | boosting | gp. Backed by scikit-learn (boosting prefers XGBoost → LightGBM → CatBoost → sklearn).
family: tabular
method: linear
data:
path: data.csv
target: y
eval:
metric: mse
min_score: null
recipe.yaml · llm / lora / qlora
family: llm with method llm | lora | qlora. model: is required (hub id or path). size: is a label only. QLoRA needs CUDA + bitsandbytes.
family: llm
method: lora
model: meta-llama/Llama-3.2-1B-Instruct
objective: lora
rank: 16
alpha: 32
steps: 100
lr: 2.0e-4
data:
path: data.jsonl
prompt: prompt
completion: completion
eval:
metric: loss
Objectives that work: next-token, sft, full-ft, lora/qlora, fim, mlm, span, continued-pretrain. Fail closed (not faked): formats, speculative, paged_kv, objective: mtp.
recipe.yaml · transformer
method: transformer with arch encoder | decoder | encoder-decoder. model: required.
method: transformer
model: bert-base-uncased
arch: encoder
steps: 50
data:
path: data.csv
text: text
target: label
eval:
metric: accuracy
recipe.yaml · guard
Opt-in fail-closed watches mid-train. See Metrics & guard.
guard:
safety: true # NaN / blow-up
leak: true # train vs evals overlap