Update HF dataset cards

[prrao87]

Rewrites every published dataset card for Hugging in this repo to a canonical six-section body: Search, Curate, Evolve, Train, Versioning, Materialize a subset.

What changes in each card

• Consistent top-of-card spine. YAML frontmatter → title + summary → key features → splits → schema → pre-built indices → "Why Lance?" → loaders (datasets.load_dataset → LanceDB → pylance) → local-download tip.
• Single LanceDB docs link in the LanceDB loader intro — exactly once per card, never repeated.
• One pylance entry-point snippet at the top of every card; the six body sections all use LanceDB.
• Materialization discipline. No unbounded to_table / to_pandas / to_arrow / to_pylist. In-memory terminals are bounded by .limit(k).to_list(). .to_batches() is reserved for the Materialize-a-subset section, piping a filtered scan straight into a new local table.
• Hub paths in read-only sections, local paths in mutating sections — Search, Curate, Versioning lookups all run against hf://datasets/...; Evolve, tag creation, and local-index builds run against .//data with a one-line pointer back to the download tip.
• SQL-first idioms for filters (.where("...")) and column derivation (add_columns({"col": "SQL_EXPRESSION"})).
• Real column names throughout, so every snippet is runnable copy-paste against the published dataset.
• Modality-aware Train sections. Image, image+text, annotation, and text-only cards use the standard Permutation.identity(tbl).select_columns([...]) pattern wrapped in a PyTorch DataLoader. Video and audio cards frame Train as a pre-extract-once / train-on-derived-features workflow, with a note that the inline blob column earns its place around the training process (inspection, sampling, evaluation, the pre-extraction pass) rather than on the per-batch hot path.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
lancedb-bcbb4faf	🟢 Ready	View Preview	May 15, 2026, 6:50 PM