Рецепты обучения нейросетей на PyTorch

★ 7.7 · learning

ml-training-recipes is a Claude Code skill that delivers battle-tested PyTorch training recipes across every major domain — LLMs, computer vision, diffusion models, medical imaging, drug discovery, protein modeling, and genomics. It ships six reference files covering training loops, optimizer configuration (AdamW and Muon with per-parameter-group settings), mixed-precision training with bfloat16, learning rate scheduling, Chinchilla scaling-law tables, and an autonomous experiment loop inspired by Karpathy's autoresearch. Architecture selection is guided by data type and dataset size, from ESM-2 embeddings for small protein datasets to training ViT from scratch at scale. The skill is built for engineers debugging loss spikes or OOM errors, fine-tuning LLaMA/GPT with LoRA, or maximizing GPU throughput on Ampere hardware. Requires torch>=2.0.0.

• #pytorch-training
• #llm-fine-tuning
• #optimization-recipes
• #vision-models
• #debugging
• #gpu-throughput
• #mixed-precision