Fashion Product Images Small — Hugging Face (ashraq)

It is the small edition of the Myntra-derived fashion catalogue that circulates widely for apparel machine learning: 44,072 products, each described by ten text or numeric attributes plus a resized photograph. Download size is 271,496,441 bytes, expanding to 546,202,015 bytes once decoded, and the whole thing sits in two Parquet shards auto-converted by the Hugging Face Hub under its data/ directory. The card carries the size category 10K–100K with modality tags for image and text.

What makes this copy the convenient one is completeness. A sibling Hub upload (ceyda/fashion-products-small) ships only 42,700 rows and resizes images to a 512-pixel maximum dimension; this version keeps all 44,072 rows intact. Downstream artefacts — including the FashionGPT Space and item-search demos — build on it, and engagement sits at 44 likes and roughly 1,658 downloads.

The tabular half reproduces the Kaggle styles.csv schema unchanged: id joins back to the original release, gender through articleType form a strict three-tier category hierarchy, and baseColour, season and usage add merchandising context that most hand-assembled catalogues leave out. productDisplayName is free text but tightly patterned — colour words appear inside it, so it doubles as weak supervision for colour extraction.

Because the categories are closed sets rather than open strings, the corpus supports label-space experiments directly: train a classifier over the 141 article types, then evaluate whether predicted labels agree with subCategory constraints. Only 86 of the 1,744 datasets in Datadory's catalog ship Parquet, so this record loads into Polars or Dask without a conversion step most peers require.

Static. The mirror was last modified on 2022-11-01, and the underlying assortment predates that — product years run roughly 2009–2017. Static snapshots are not rare in this catalog: 181 of 1,744 datasets never update, and research corpora are where they cluster. Plan around a fixed corpus rather than an incremental feed, and pair it with a living complement when recency matters — Lyst's catalogue spans more than 27,000 brands, Zalando's covers 7,000+ brands across 29 markets, and both sit in the same industry brief. For transaction-level demand signal instead of catalogue structure, H&M Group's release carries 31,788,324 purchases across 105,542 articles.

Six of Datadory's eight personas clear the relevance bar for this record, ranked here by fit:

1. Data Scientists & ML Engineers (relevance 2/3) — stream all 44,072 rows for embedding and retrieval experiments without unpacking archives. Industry-specific patterns live on our data scientists use cases page.
2. Developers & Data-Product Builders (relevance 2/3) — load thumbnails straight into an app pipeline with a documented feature schema, no archive handling required; see developers builders use cases.
3. Market Researchers & Consultants (relevance 1/3) — inspect Indian online apparel assortment structure when the original Kaggle source is awkward to fetch.
4. Competitive Intelligence & Product Teams (relevance 1/3) — review the attribute vocabulary a top marketplace assigns, to calibrate your own product-data checklist; more on competitive intel product teams use cases.
5. E-commerce Operators (relevance 1/3) — reference realistic attribute completeness across colour, usage and season when defining intake standards for your own catalogue.
6. Journalists, Academics & Students (relevance 1/3) — demonstrate dataset-provenance risk using a popular mirror that declares no licence anywhere on its card.

It wins on loading friction and loses on scale and rights clarity. Within its own family: the Kaggle original ships the same 280 MB-scale imagery behind a Kaggle account with an explicitly declared licence; the 44k Fashion Product Images Dataset adds roughly 24.8 GB of high-resolution photography; the ceyda copy trades 1,372 rows for smaller 512-pixel images. Against the wider machine-learning corpora, DeepFashion dwarfs it with 800,000+ annotated images across 50 categories and 1,000 attributes, Fashionpedia brings 46,781 images with 342,182 bounding boxes, and Fashion-MNIST remains the 70,000-image baseline benchmark. None of those load as a single flat table of 44,072 labelled SKUs with a closed attribute vocabulary — this record's defining advantage.

Datadory scores it 7 out of 10 on its rubric of field documentation, access reliability and freshness — a band shared by 296 of the 1,744 cataloged datasets, just below the catalog-wide mean of 7.81. Field definitions are marked verified: the feature schema was cross-checked against repository metadata and live row output during research. Sample data exists and is quotable — id 15970, "Turtle Check Men Navy Blue Shirt", Men/Apparel/Topwear/Navy Blue, Fall 2011.

Three things hold the score at 7. The dataset card is a stub documenting nothing beyond the feature schema; no licence is declared anywhere on the card, repo YAML or API metadata, even though the source Kaggle release carries a licence label applied to marketplace-sourced imagery — treat rights verification as part of your procurement step; and streaming mode against the auto-converted shards went untested during research. Engagement metrics suggest the repo stays healthy regardless.