Add a dataset

Choose the smallest supported pattern: custom for a standard CSV, single only for a distinct loader, or the traffic converter for values plus adjacency.

uv run tsf dataset add --name my_data --pattern custom \
  --path ./dataset/my_data/my_data.csv --target OT

Keep bytes in dataset/, loader/schema code in src/data/, the runnable preset in configs/datasets/, and the generated card in catalog/datasets/. A preset uses one dataset.path; only catalog-style datasets such as GIFT-Eval add a separate dataset.id.

Keep [dataset] limited to name, optional display/track fields, path, optional id, and [dataset.params]. Put loader-specific options in a strict registered parameter schema; reject misspellings and catch-all keyword arguments. A file loader receives a full file path, a directory loader receives a directory, and a selector loader requires both path and id.

For traffic bundles, inspect uv run tsf dataset convert-traffic --help and provide explicit inputs, splits, and windows. Then inspect and exercise the data:

uv run tsf dataset inspect --config configs/datasets/my_data.toml
uv run tsf inspect --config configs/runs/<run>.toml
uv run tsf run configs/runs/<run>.toml

Confirm train-only scaling, feature/target selection, split boundaries, and adjacency injection. Stop before overwriting an existing dataset unless replacement was requested.

tsf dataset add also regenerates the preset's canonical README card. Complete any loader-specific facts in executable config/code, regenerate with uv run python scripts/generate_resource_cards.py, then require both uv run tsf dataset show my_data and uv run tsf dataset audit to pass.