Apparel, Accessories & Luxury Goods · Hugging Face Datasets
Fashion Products Small — Hugging Face (ceyda)
Datadory delivers fashion products small Hugging Face ceyda data for apparel, accessories and luxury goods teams: 42,700 Myntra product rows with embedded 512-pixel images, filename-and-id joins back to the original photographs, and masterCategory, subCategory and gender labels across 45 subcategories. Shipped as Parquet via API, files, or your warehouse on the cadence you choose.
- Where it covers
- India — Myntra e-commerce catalogue
- How far back
- Snapshot frozen July 2022; underlying products roughly 2009–2017
- How fine
- Product level — one row per SKU with an embedded resized image
What is the Fashion Products Small dataset on Hugging Face?
Fashion Products Small is a preprocessed copy of the well-known 44k Myntra fashion catalogue, uploaded to Hugging Face under the handle ceyda and trimmed to a single train split of 42,700 examples packaged in two Parquet shards. Every product row carries an embedded image resized so no dimension exceeds 512 pixels, plus the taxonomy labels that make the set trainable straight away: masterCategory, subCategory and gender. The compressed payload is 596,574,363 bytes, unpacking to roughly 661 million bytes of Parquet, and the original Myntra photo links survive in their own column so each row still resolves to its full-resolution photograph. One candid caveat: the card describes the project as built for test purposes and its metadata is thin — precisely why we lead with samples before you commit.
What do the first rows actually look like?
Three records, exactly as the copy stores them. Each of the 42,700 rows also embeds the resized image bytes inline, so a training loop reads pixels and labels from one record with no second join; observed image widths run from 150 to 512 pixels.
filename | id | masterCategory | gender | subCategory | link (truncated)
----------|-------|----------------|--------|-------------|-------------------------------------------
15970.jpg | 15970 | Apparel | Men | Topwear | assets.myntassets.com/v1/…/7a5b82d1372a7a….jpg
39386.jpg | 39386 | Apparel | Men | Bottomwear | assets.myntassets.com/v1/…/4850873d0c4176….jpg
59263.jpg | 59263 | Accessories | Women | Watches | assets.myntassets.com/v1/…/Titan-Women-Silver-Watch_….jpgGet a sample of this dataset — request rows, the full field dictionary, or a warehouse-mounted extract and we route it the same day.
What fields are in the dataset?
Seven columns per row, all verified against the served schema:
| Field | Type | Definition | Example |
|---|---|---|---|
filename | string | Original image filename inherited from the source catalogue; the join key back to any fuller archive. | 15970.jpg |
id | string | Numeric product identifier stored as a string, stable across sibling copies of the catalogue. | 15970 |
masterCategory | string | Top-level product category, 7 distinct values. | Apparel |
subCategory | string | Second-level category, 45 distinct values. | Topwear |
gender | string | Target gender segment, 5 distinct values. | Men |
link | string | Product-photo link on the Myntra asset host, preserved from the source catalogue so any row resolves to its full-resolution photograph. | assets.myntassets.com/v1/…/7a5b82d1….jpg |
image | image | Embedded photograph resized to a maximum dimension of 512 pixels; widths observed between 150 and 512. | — |
Observed masterCategory values include Apparel, Accessories, Footwear and Personal Care; subCategory runs 45 deep, from Topwear, Bottomwear and Shoes to Watches and Bags; gender spans Men, Women, Boys, Girls and Unisex. What this copy deliberately drops, relative to the Kaggle original and the ashraq mirror on the same hub, is articleType, baseColour, season, usage and productDisplayName. Additional fields on request: we can append those attribute columns from the fuller catalogue copies to your delivery, joined on the shared string id.
Where does the coverage sit?
- Geography: India — the Myntra e-commerce catalogue the set was drawn from.
- Temporal: a static snapshot frozen in July 2022; the underlying products date from roughly 2009–2017.
- Granularity: product level — one row per SKU, each with its embedded resized image.
How is Fashion Products Small delivered?
API, files, or your warehouse. Daily, weekly, or hourly.
Who uses this dataset — and for what?
- ML and computer-vision teams treat it as a ready-made training corpus for garment category classifiers and visual-similarity retrieval: labels and pixels arrive in the same row, and 42,700 examples is enough headroom for fine-tuning without a download marathon.
- E-commerce operators benchmark their own category trees against a live retailer taxonomy — 7 master categories fanning into 45 subcategories is a compact reference hierarchy.
- Competitive intelligence and product teams read assortment mix by gender segment and subcategory, using the preserved photo links to spot-check how products were originally presented.
- Developers and builders prototype search ranking and cold-start recommendations on the taxonomy backbone before committing to heavier imagery pipelines.
If you are weighing it against the fuller mirrors, the honest trade is width of attributes versus weight: this copy trades colour, season and usage labels for a payload under 600 MB compressed.
Which datasets pair well with it?
- Fashion Product Images Small — Hugging Face (ashraq) — the 44,072-row sibling mirror that keeps the articleType, baseColour, season, usage and productDisplayName columns this copy drops.
- Fashion Product Images Dataset (44k) — Kaggle — the full-resolution parent catalogue, roughly 24.8 GB across 44,000+ products.
- Fashionpedia — Detection Datasets on Hugging Face — 46,781 web-sourced fashion images with 342,182 bounding boxes and 294 attributes when you need localization, not just classification.
- Fashion-MNIST — Zalando Research (GitHub) — 70,000 grayscale thumbnails that make a fast sanity baseline before real imagery enters the loop.
- Apparel, Accessories & Luxury Goods hub — all 23 catalogued datasets for the industry, trade series and taxonomies included.
Questions buyers ask
How many rows does Fashion Products Small contain?
42,700 product rows in a single train split, packaged as two Parquet shards totalling 596,574,363 bytes compressed and about 661 million bytes unpacked. Each row pairs three taxonomy labels with an embedded image whose longest side is capped at 512 pixels. The fuller Kaggle original claims roughly 44,000 products, so plan cross-copy joins accordingly.
Which labels ship with each product row?
Three categorical layers: masterCategory with seven values such as Apparel, Accessories, Footwear and Personal Care; subCategory with 45 values running from Topwear and Bottomwear to Watches, Bags and Shoes; and gender spanning Men, Women, Boys, Girls and Unisex. Filename, string product id and the preserved Myntra photo link complete every record.
Why does this copy have fewer rows than the Kaggle original?
Nobody has documented the delta. This copy holds 42,700 rows, the ashraq mirror on the same hub holds 44,072, and the Kaggle upload claims about 44,000. Treat the missing few thousand as unverified until you diff identifiers — the string id column makes that check cheap, and we can run it as part of a sample request.
Is the image resolution good enough for model training?
For taxonomy and retrieval work, yes: images are resized to a maximum dimension of 512 pixels, with widths observed between 150 and 512. For fabric-level detail or print-quality reproduction, resolve the preserved Myntra photo link instead of relying on the embedded bytes — that column bridges back to full-resolution photography.
Can the dropped colour, season and usage labels be added back?
Yes. This copy omits articleType, baseColour, season, usage and productDisplayName, which the fuller catalogue copies retain. Request them and we append those columns to your delivery, joined on the shared string id, keeping this copy's lighter footprint while recovering the complete attribute set.
See the rows before you pay anything.
Name this dataset and we send real records from it — scoped to the fields you asked for.