Hugging Face Skills

Hook up Claude Code to Hugging Face so you can load models and datasets, inspect weights, pull model card metadata, and spawn Spaces without leaving your editor. You can call endpoints, switch a model to inference mode, or download checkpoints. The interface exposes dataset tables and sample rows so you can preview examples and schema before you commit to training.

Use the TRL stack on cloud GPUs to train and fine-tune language and RLHF-style models: start jobs, watch logs, cancel runs, and download checkpoints. Create and manage datasets from SQL queries, run preprocessing steps, and map tokenization across splits. You can run evaluation scripts, collect metrics, and compare runs side-by-side.

Track experiments by recording hyperparameters, artifacts, and metrics, then pull past runs for analysis. Publish a Space, push model cards, or upload datasets and metadata to Hugging Face from Claude Code. The tool reads error traces and suggests next steps so you can iterate faster.

In practice you can run a quick fine-tune, evaluate against a test suite, and deploy the best model to a Space without switching browser tabs — useful when you need to ship a bugfix model for content moderation or prototype an NLU update during a sprint.