Model Registry: HuggingFace Hub Privat Repo + MLflow + S3 Layout + Versioning

Production'da 50+ fine-tuned model versiyonu nasıl yönetilir? HuggingFace Hub privat repo + MLflow Model Registry + S3 (parça-parça artifact) hybrid. Versioning convention (semantic versioning + lineage), tags (\`production\`, \`canary\`, \`archive\`), retention policy (eski versiyon ne zaman silinir?). Cookbook'un model card şablonu (LoRA adapter + base + recipe).

Şükrü Yusuf KAYA

28 dakikalık okuma

14.05.2026

İleri

Model Registry: HuggingFace Hub Privat Repo + MLflow + S3 Layout + Versioning

1. Cookbook Model Registry Hiyerarşisi#

HuggingFace Hub (privat repo): kompanyam/llm-models
  ├── llama-3.1-8b-tr-instruct-v1.0/        # Stable baseline
  ├── llama-3.1-8b-tr-instruct-v1.1/        # Minor improvement
  ├── llama-3.1-8b-tr-instruct-v2.0/        # Major retrain
  ├── llama-3.1-8b-tr-customer-support-v1.0/  # Domain variant
  └── ...

Her repo içinde:
  - adapter_model.safetensors             # LoRA weights
  - adapter_config.json                    # PEFT config
  - tokenizer.json / tokenizer_config.json
  - README.md                              # model card (zorunlu)
  - eval_results.json                      # benchmark sonuçlar
  - training_config.yaml                   # reproducible
  - WANDB_RUN_URL                          # full training telemetry

Lineage triple (Part 0 Ders 0.5):