zelda-model-manager

name	zelda-model-manager
description	Manage Zelda/ALTTP/Oracle model training, datasets, evals, registry updates, and deployment artifacts (Nayru/Din/Farore/Veran/Sahasrahla/IQuest). Use when planning or running training runs, curating ASM datasets, selecting base models, evaluating outputs, or converting/serving GGUF or MLX models.

Zelda Model Manager

Scope

• Manage Zelda and ASM model lifecycle: dataset inventory, training runs, evals, registry, and deployment artifacts.

Workflow

1. Confirm the target model role and naming.
   • Use ~/src/docs/NAMING_CONVENTIONS.md and ~/src/lab/afs-scawful/docs/MODEL_PORTFOLIO.md.
   • Keep hostnames as SSH aliases (medical-mechanica, halext-nj) instead of IPs.
2. Locate datasets and scripts before deciding on a run.
   • Read ~/src/training/INDEX.md and ~/src/training/README.md for dataset paths and scripts.
   • For large runs, consult ~/src/training/docs/IQUEST_40B.md.
   • For smaller Zelda plans, consult ~/src/lab/afs-scawful/docs/ZELDA_16B_TRAINING_PLAN.md.
3. Choose base model and hardware based on tool-calling needs.
   • Prefer Qwen 2.5 Coder for tool calling and ASM workflows.
   • Follow ~/src/training/docs/MODEL_SELECTION_AND_PRACTICES.md.
4. Run QA before training.
   • Use dataset QA and registry updates described in ~/src/lab/afs-scawful/docs/ZELDA_16B_TRAINING_PLAN.md.
   • Ensure AFS dataset index is current (python -m afs_scawful datasets index).
5. Monitor training and evaluate.
   • Use eval packs in ~/src/training/evals/.
   • Track ASAR pass rate for ASM validity.
6. Register and deploy artifacts.
   • Use the AFS registry (~/src/lab/afs-scawful/config/chat_registry.toml) to define personas, ports, and parameters.
   • Use ~/src/tools/model-mgr/model-mgr for GGUF/MLX conversion and Ollama imports.
   • Test deployments using python3 ~/src/lab/afs/lmstudio_client.py (checks health and ports).

Commands to reuse

• model-mgr list and model-mgr info <model> for inventory.
• model-mgr convert <model> --quantize q4km for GGUF.
• model-mgr mlx-convert <model> --hf-path <path> for MLX exports.

Knowledge References

Consult the global knowledge base at ~/.context/knowledge/models/ for background:

• Model portfolio & status: models/portfolio.md
• Training pipeline architecture: models/training-pipeline.md
• Dataset catalog: models/datasets.md
• GGUF conversion & deployment: models/infrastructure.md
• Step-by-step workflows: models/workflows.md
• Serving & routing: models/serving.md

References

• Read references/sources.md for source paths and anchors.