SLM Packages

An SLM package is the unit of distribution in SLMCortex.

It wraps a trained LoRA adapter with the metadata needed to validate, route, compose, and reproduce the package later. Packaging records provenance, fingerprints, protected input snapshots, evaluation data, and optional composition metadata.

Typical package layout:

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

examples.jsonl is optional. A package uses one adapter weight format: MLX packages use adapter/adapters.safetensors; GGUF packages use adapter/adapter.gguf.

Create a package from an existing adapter:

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 before composition:

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