Aulafy
Courses/AI security and evaluation/Model and dataset supply chain

Model and dataset supply chain

Your AI app inherits trust from everything it downloads: model, tokenizer, dataset, LoRA, npm package, Docker image, workflow, script, and extension.

  • Create a provenance record for models and datasets.
  • Avoid uncontrolled weights, scripts, and dependencies.
  • Treat license, security, and quality as separate decisions.

Provenance record

TerminalCode
modelo:
  nombre: "qwen3-8b"
  fuente: "repositorio oficial"
  version_o_commit: "..."
  licencia: "revisada"
  formato: "gguf | safetensors | onnx"
  hash: "sha256:..."
  requiere_trust_remote_code: false
  datos_entrenamiento_conocidos: parcial
  uso_permitido: "educativo | comercial | revisar"
  fecha_revision: "2026-07-03"

Simple rules

  • Prefer formats that do not execute code when loading weights.
  • Avoid trust_remote_code unless you review exactly what it executes.
  • Store hashes of weights used in production.
  • Do not mix private and public datasets without logging.
  • Review model, dataset, and adapter licenses separately.
TerminalCode
sha256sum modelo.gguf > checksums.txt
pip freeze > requirements.lock.txt
npm ls --depth=0 > npm-deps.txt
docker image inspect qdrant/qdrant:latest > docker-qdrant.json