Add ML Training Recipes skill

[dailycafi]

Summary

Adds a comprehensive ML training recipes skill to 10-optimization/, covering battle-tested PyTorch training patterns across all domains.

What's included

• SKILL.md (319 lines): Training loops, optimizer config, LR scheduling, mixed precision, debugging checklist, experiment management
• 6 reference files (96KB total):
  • architecture.md — Transformer/LLM architecture patterns, weight init
  • optimizers.md — Muon, AdamW hybrid, per-group LR, compiled steps
  • domain-specific.md — Vision, diffusion, data loading, architecture tables
  • scaling-and-selection.md — Chinchilla scaling, compute budgets, DGX Spark
  • biomedical.md — Drug discovery, protein LMs, medical imaging, genomics, spatial omics, clinical NLP
  • experiment-loop.md — Autonomous experiment loop (autoresearch-style keep/discard/revert)

Key differentiators from existing skills

This skill fills gaps not covered by existing optimization skills (which focus on quantization/inference):

• Muon optimizer (polar express orthogonalization) — cutting-edge, not in any existing skill
• Karpathy's debugging checklist — systematic training diagnosis
• Autonomous experiment loop — fixed time-budget, keep/discard methodology
• DGX Spark bandwidth-limited optimization — specialized hardware patterns
• Biomedical ML — molecular GNNs, ESM-2, MONAI, nnU-Net, spatial omics
• Per-parameter-group optimizer config — different optimizers for embeddings vs matrices

Sources

• Karpathy autoresearch / nanochat
• Muon Optimizer
• Google Deep Learning Tuning Playbook
• Chinchilla scaling laws (Hoffmann et al., 2022)
• μP (Yang et al., 2022)

Complementary to existing skills

This skill focuses on training methodology while existing 10-optimization skills focus on inference optimization:

• No overlap with Flash Attention, bitsandbytes, GPTQ, AWQ, GGUF, HQQ
• Complements 08-distributed-training (covers single-GPU → multi-GPU bridge)
• Complements 03-fine-tuning (provides the training recipe framework)

Quality checklist

• YAML frontmatter with all required fields
• SKILL.md: 319 lines (within 200-300 target)
• Progressive disclosure (SKILL.md overview → reference files for depth)
• Code examples with language detection
• Debugging checklist with solutions
• marketplace.json updated and validated

[zechenzhangAGI]

exciting direction on autoresearch. thanks for the proposal! will take a look