raghu619/shopify_category_lite

Shopify Product Category (curated lite)

A curated, balanced subset of Shopify/product-catalogue for product categorisation from text: read a product's title and description, predict its category.

Built as a learning project applying data-curation practice (funnel tracking, label design, leakage checks, deduplication, rebalancing, text hygiene) to a real public catalogue.

Code and write-up: raghu619/catalog-curation (curation notebook, rewrite notebook).

Task and columns

Column	Meaning
text	Model input: product title + description (description truncated to 1000 chars)
summary	The same product rewritten by an LLM into a standard 4-line English format (see below)
label	Target: level-2 category, e.g. Home & Garden > Decor (141 classes)
full_category	The original full category path, for reference

Splits: 20,027 train / 2,498 validation / 2,497 test, stratified by label.

Curation funnel

Step	Products
raw_train	38,631
raw_test	9,658
labelled_train	38,511
labelled_test	9,624
merged	48,135
no_conflicting_titles	48,106
deduplicated	48,074
classes_20plus	47,550
capped	25,023
train	20,027
validation	2,498
test	2,497

Decisions

• Label granularity: level 2. Level 1 is too broad to be useful; at level 3, 4 out of 5 classes had fewer than 20 products.
• Removed non-product labels (Uncategorized, Gift Cards, Bundles, Product Add-Ons) and 1-level paths.
• The source's potential_product_categories column is excluded. It contains the correct full path in 100% of products, and in 23.5% of products every candidate points to the correct level-2 class. Using it turns name the category into pick from a list.
• Dropped 8 titles that appeared with conflicting labels (29 products; placeholder items such as example product).
• Deduplicated on normalised title + description after merging the source splits: 32 duplicate products removed.
• Rebalanced: classes with fewer than 20 products dropped; classes capped at 500 (random sample).
• Text hygiene: removed emails (283 products), URLs (300) and phone numbers (74). A first, looser phone rule matched 1,813 products, mostly sizes, years and IDs, and was replaced with a keyword-anchored rule.
• Languages kept as-is: mostly English, with Spanish, French, German, Italian, Portuguese and Japanese. (A Day 1 word heuristic estimated ~9% Spanish; it could not see other languages.)
• Duplicates re-checked after cleaning: one product became identical to a train product only after text hygiene; it was removed from test. No text appears in more than one split.

LLM rewrite (summary column)

Every product's text was rewritten by gpt-4.1-nano (OpenAI Batch API, temperature 0) into:

Title: ...
Brand: ... (or Unknown)
Description: one sentence (or Unknown)
Details: one sentence on key features (or Unknown)

Prompt rules (prompt v2): no category line (the category is the label), never invent facts, write Unknown when information is missing, copy numbers exactly (decimal commas like 12,5 cm), remove IDs, URLs and contact details, translate to English. The 4-line structure is then enforced in code (clean_summary): first block only, missing Title taken from the listing's first line.

Quality, all 25,022 products:

Check	Result
Bad format	0
Invented numbers flagged	358 (1.43%); in 20 hand-labelled cases ~25% were real (rest were unit/year conversions), so roughly 0.35%
Description = Unknown	391 (1.56%); 73 of these had real descriptions (mostly books and plants), a known limitation
Median length	575 -> 260 characters
Cost	$0.77 for the full run

Prompt history: v1 invented facts (a 12,5 cm measurement became "12 sheets") and missed obvious brands. v2 fixed both. v3 was tested on all products and rejected: Description = Unknown rose from 1.56% to 2.50% and flagged invented numbers from 1.43% to 1.73%.

Train/serve note: if you train on summary, rewrite new products with the same prompt and clean_summary before predicting.

Known limitations

• Capping classes discards about half the usable products and makes test scores reflect a balanced catalogue, not the real mix.
• Some source labels are genuinely ambiguous (e.g. an "Airplane Piggy Bank": toy or decor?). This caps achievable accuracy.
• The language estimate is a word-count heuristic, not a language detector.
• About 7% of source descriptions are empty; those products rely on the title alone.

Source and license

Derived from Shopify/product-catalogue, released under Apache-2.0. This derivative is released under the same license. Images from the source are not included.