// Use when creating or revising an HF_DATASET_CARD.md file in this repo. Defines the six-section card structure (Search, Curate, Evolve, Train, Versioning, Materialize a subset), the writing voice, and the example style rules for Lance-format datasets published to the HuggingFace Hub.
| name | hf-dataset-card-writer |
| description | Use when creating or revising an HF_DATASET_CARD.md file in this repo. Defines the six-section card structure (Search, Curate, Evolve, Train, Versioning, Materialize a subset), the writing voice, and the example style rules for Lance-format datasets published to the HuggingFace Hub. |
This skill governs the public-facing HF_DATASET_CARD.md published with each Lance-format dataset on the HuggingFace Hub. Treat the card as user-facing documentation for ML engineers, researchers, and agent/app builders consuming the dataset. Do not copy maintainer workflows (data generation, conversion internals, upload mechanics) into the card unless explicitly requested.
Examples should make the value of Lance self-evident — show, don't tell. Never name or compare to other formats (Parquet, WebDataset, image folders); demonstrate the strengths and let the reader draw their own conclusion. Do not label sections with persona tags ("for agents", "for researchers", etc.) — the same section often applies to multiple personas, and readers should not have to self-identify before reading.
Prose should be informative and educational, written in full, flowing sentences that explain what the code does and why it matters. Avoid the staccato style of fragments stitched together with em-dashes, parenthetical asides, and bullet-like one-liners; readers should be able to learn the concept from the prose alone, with the code as confirmation. Each section's introduction should set up the problem the snippet addresses, name the relevant Lance mechanism, and explain the consequence of using it. Short follow-up paragraphs after a snippet should expand on a non-obvious property of the result rather than restate what the code did.
Bulleted lists, terse callouts, and tables remain appropriate for the inherently list-shaped parts of the card (key features, schema, indices, tips). The prose-style rule applies to the body of each of the six sections below. Write in an informative, educational manner without being verbose; use em-dashes sparingly.
The top of every card contains, in order:
task_categories values are validated against HF's official allowlist — do not put lance there or the Hub will surface a "YAML Metadata Warning" on the card. Put lance in tags instead (which is free-form).datasets.load_dataset, lancedb.connect, then lance.dataset. LanceDB comes before pylance because it is the higher-level interface most users interact with; pylance is shown below as the format-level handle for readers who want to inspect or operate on dataset internals. The LanceDB loader's intro sentence carries the single LanceDB-docs link for the card: LanceDB is the embedded retrieval library built on top of the Lance format ([docs](https://lancedb.com/docs)), and is the interface most users interact with. That is the only place the docs URL appears — do not add it elsewhere.hf download commandInsert verbatim between the Pre-built indices section and the datasets.load_dataset loader. Do not customize per card:
## Why Lance?
1. **Blazing Fast Random Access**: Optimized for fetching scattered rows, making it ideal for random sampling, real-time ML serving, and interactive applications without performance degradation.
2. **Native Multimodal Support**: Store text, embeddings, and other data types together in a single file. Large binary objects are loaded lazily, and vectors are optimized for fast similarity search.
3. **Native Index Support**: Lance comes with fast, on-disk, scalable vector and FTS indexes that sit right alongside the dataset on the Hub, so you can share not only your data but also your embeddings and indexes without your users needing to recompute them.
4. **Efficient Data Evolution**: Add new columns and backfill data without rewriting the entire dataset. This is perfect for evolving ML features, adding new embeddings, or introducing moderation tags over time.
5. **Versatile Querying**: Supports combining vector similarity search, full-text search, and SQL-style filtering in a single query, accelerated by on-disk indexes.
6. **Data Versioning**: Every mutation commits a new version; previous versions remain intact on disk. Tags pin a snapshot by name, so retrieval systems and training runs can reproduce against an exact slice of history.
Keep meta/info.json in the dataset card because users rely on it for global metadata and feature definitions. Explain the purpose of each table in plain language (what it is for, when to use it), not only terse schema bullets.
Every card then carries six body sections in this order:
Show a regular vector search using the dataset's bundled vector index. If the dataset also ships an FTS index, follow up with a hybrid search (vector + FTS) example. Keep this section focused on retrieval mechanics; versioning lives in section 5. All six sections appear in every card, even when one feels forced for the modality.
Ad-hoc filtering and slicing that produces a small, explicit candidate set for the next experiment. Always bounded by .limit(...). Output is framed as a row list that feeds into Evolve or Train.
Add features without rewriting the dataset. Prefer tbl.add_columns({"new_col": "SQL_EXPRESSION"}) for in-card examples: a SQL expression that derives the new column from existing ones is a one-liner with no decorators, no Arrow construction, and no fake placeholder values. Pair it with tbl.merge(other, on="key") to attach offline labels or predictions from an external table, if required. Mention that original columns and indices are untouched.
Feed a PyTorch loop with column projection. Use lancedb.permutation.Permutation.identity(tbl).select_columns([...]) wrapped in torch.utils.data.DataLoader. select_columns(...) is the lever: only the projected columns are read per epoch, so columns added in Evolve cost nothing until you opt in. The card should open with db = lancedb.connect(...) / tbl = db.open_table(...) (same handle used by Search, Curate, and the rest of the card) so readers see one consistent entry point through LanceDB.
Video and audio datasets need a different Train framing. When each row carries a full clip in a blob-encoded column (MP4 bytes, WAV bytes, etc.), a common practitioner pattern is to pre-extract decoded frames or audio samples once into a derived Lance dataset (one row per training-ready window) and train against that derived dataset using the regular Permutation.identity(tbl).select_columns(...) pattern. Other workflows project the blob column directly through select_columns(...) and decode at the batch boundary, or skip the media entirely and train on cached embeddings; the right shape is workload-specific. The card's Train section for these modalities should mention these options in a brief, neutral opener — note that take_blobs is the mechanism for subsetting clip bytes on demand inside a one-time pre-extraction pass — and then show only the actual training-loop snippet(s) (e.g., a Permutation against a notional pre-extracted frames table, and if the dataset ships precomputed embeddings, a short variant that projects those from the source table). Do not write out the pre-extraction code itself; window length, sampling rate, resize policy, and on-disk frame format are workload-specific, and showing one concrete recipe risks teaching it as the only one. The inline blob storage on the original dataset earns its place for transport, random-access inspection, curation, and the pre-extraction step itself; whether it also feeds the dataloader is the reader's call.
Show how to list versions / tags and open a pinned snapshot (db.open_table(..., version=...) or via tags). Cover both use cases without labeling personas: looking up past data versions for added context at serving time, and running experiments with different models against different data versions. This is where the reproducibility and governance story lives.
Stream a filtered query from the Hub-mounted table into a new local LanceDB table using tbl.search().where(...).select([...]).to_batches() piped into db.create_table(name, batches). Explain that reads from the Hub are already lazy, that mutations and fast training need a writable local store, and that this pattern transfers only the projected columns and matching row groups without ever fully materializing in Python memory. State explicitly that the Evolve, Train, and Versioning snippets above can target the local subset by swapping the path. This is the least essential of the six sections — it sits at the end as a reference pattern that earlier sections can point back to when readers need a writable copy.
to_table(), to_pylist() (without a bounded take / limit), to_arrow(), or to_pandas() on a scanner or search in examples — readers will copy these against million-row datasets and OOM or stall. Bounded ds.take([id1, id2, ...]) is fine because it's index-driven.db.open_table(...) then tbl.search(...).where(...).select([...]).limit(k).to_list(). .limit(k) is mandatory before any in-memory terminal (.to_list(), .to_arrow(), .to_pandas())..limit() rule. .to_batches() returns a pa.RecordBatchReader and does not accumulate rows in Python memory, so it is safe (and expected) without a .limit() when the goal is to pipe a filtered query into another sink — typically db.create_table(name, batches) in the Materialize-a-subset section.lance.dataset(...) + count_rows() + list_indices() block) so readers see the low-level handle. Use LanceDB everywhere else — Search, Curate, Evolve, Train, Versioning, and Materialize-a-subset. Drop down to pylance inside a body section only when LanceDB does not expose the operation (e.g., ds.take_blobs(...) for lazy blob handles on a video or audio dataset).hf://datasets/<org>/<repo>/data (LanceDB) or hf://datasets/<org>/<repo>/data/<split>.lance (pylance) so a reader can run the code without downloading anything first. Use a local path (e.g., ./<repo>/data) only in snippets that mutate the dataset — add_columns, merge, create_index, tags.create — because the Hub mount is read-only. When introducing a local path, add a one-line note explaining why and pointing back to the hf download tip (or the Materialize-a-subset section) near the top or bottom of the card..where(...) and SQL derivation expressions in add_columns({col: sql_expr}) are the universal idiom and require no extra imports or decorators. Lean on SQL anywhere the operation can be expressed declaratively.The current canonical example of a card written to this spec is laion-1M/HF_DATASET_CARD.md. Read it before writing or revising another card.