Recipe schema

Recipe schema#

A recipe is a YAML document with common fields used by the CLI and a engine specific engine_args mapping consumed by the selected engine.

Common structure#

The common top-level sections are:

  • metadata: recipe identity and compatibility requirements.

  • defaults: training defaults passed through the normalized training request. The fields and values may be training-method-specific; the recipe’s explicit metadata.method determines how they are interpreted.

  • engine_args: engine-specific configuration. Its fields are defined by the selected engine reference.

metadata:
  recipeId: aimfttk/sft/google-gemma-3-27b-it/lora/1gpu/default
  status: preview
  engine: aimfttk
  modelId: google/gemma-3-27b-it
  aimId: google/gemma-3-27b-it
  method: SFT
  isAdapter: true
  acceleratorCount: 1
  compatibleAccelerators: ["74a1", "74a9"]

defaults:
  epochs: 1
  learning_rate: 0.000005
  per_accelerator_batch_size: 1

engine_args:
  distributed_type: auto-deepspeed-stage1

Engine-specific fields#

The CLI validates and forwards engine_args to the adapter selected by metadata.engine. Avoid using engine-specific fields that are equivalent to the default parameters available in the defaults section.

See the AIM Fine-tuning Toolkit engine reference for the currently supported aimfttk fields.

Contents