Hugging Face Transformers AI Skill Guide (Claude)

Overview & Engine Architecture

Transformers is the Hub-facing stack for configs, tokenizers, and model classes (Auto* / task-specific heads) plus high-level pipeline helpers. Text becomes input_ids / attention_mask (and optional chat control tokens via chat templates); the model runs on PyTorch (primary), with optional TF/Flax. Agents pin revision, prefer safetensors, separate train vs eval + inference_mode, and treat Hub downloads as supply-chain inputs—not floating main in production.