Skip to main content

SLM Package Contract

An SLM package is the reusable artifact format in SLMCortex.

It wraps a trained LoRA adapter with the metadata needed to validate, route, compose, and reproduce it later.

Required Package Files​

  • slm.yaml
  • metadata.json
  • training_config.json
  • eval.json
  • adapter weights under adapter/

Optional files include README.md and examples.jsonl.

Typical Layout​

slms/python_slm/
├── adapter/
│   ├── adapters.safetensors
│   ├── adapter.gguf
│   └── adapter_config.json
├── slm.yaml
├── README.md
├── eval.json
├── training_config.json
├── metadata.json
└── examples.jsonl

A real package uses one adapter weight format, not both:

  • MLX packages use adapter/adapters.safetensors
  • GGUF packages use adapter/adapter.gguf

What The Metadata Is For​

The package contract preserves:

  • routing hints
  • compatibility declarations
  • dataset provenance
  • checksums
  • protected input snapshots

That is what lets Composer and Runtime Core treat a package as a trustworthy unit instead of a loose folder of weights.

Minimal Composition Metadata​

Self-describing packages can embed composition metadata directly:

composition:
  capabilities:
    allowed_task_types: [debugging]
  activation:
    default_route_type: adapter
    scope: task
  compatibility:
    compatible_slms: []
    incompatible_slms: []
  routing:
    tasks: {}

Required fields:

  • composition.capabilities.allowed_task_types
  • composition.activation.default_route_type
  • composition.activation.scope

Validation Rules​

  • required files must exist
  • metadata.json must record deterministic per-file checksums
  • protected inputs must be recorded and unchanged
  • if composition metadata exists, slm.yaml and metadata.json must agree

Create a package:

slmcortex package-slm \
  --slm-id python_slm \
  --name "Python Slm" \
  --adapter-dir artifacts/adapters/python_slm \
  --train-dataset data/train.jsonl \
  --eval-dataset data/eval.jsonl \
  --eval-summary tests/fixtures/slmcortex_demo/eval-summary.json \
  --output slms/python_slm

Validate it:

slmcortex validate-slm-package --path slms/python_slm

Compose it:

slmcortex compose-slms \
  --slms slms/python_slm,slms/debugging_slm \
  --output runtime/debugging_bundle