lxe/receipt-line-items-taxonomy

Receipt Line Items with Taxonomy

Printed US receipt line items ("GV PPR TWL 6RL", "1 Bfast English") paired with a readable product name ("Great Value Paper Towels", "English Breakfast") and one of 39 spending categories (taxonomy v2, taxonomy.json). It is the permissively licensed part of the training data behind the item namer in the Receipt Kitten iOS app. The Open Food Facts-derived part is a separate ODbL dataset: lxe/receipt-household-lines-odbl.

config	rows	what
real_lines	7,420 lines from 1,530 receipts	item lines of real receipts from openly licensed datasets, labelled by an LLM teacher
synthetic_lines	102,932	synthetic POS-style labels generated from product names (USDA Branded Foods), hand-written lexicons and real-receipt seeds, labelled by the teacher
household_synth	38,692	teacher-generated household, home, pet, beauty and hardware items for 21 US chains, with 2 printed-label variants each

Fields

Common: store, store_type (grocery, retail, restaurant, cafe, pharmacy, convenience, big-box, gas/fuel, other), label (as printed), name (clean name: brand if printed or implied, product, key variant; no sizes; at most ~6 words), category (taxonomy-v2 leaf), top_level, confidence (teacher's own: high/medium/low).

• real_lines: source, source_id (id in the source dataset), source_license, line_index, amount, section (heading printed above the item, when any), teacher_pass (1 = first pass; 2 = re-labelled by a thinking-high second pass because pass 1 was medium/low), split (train/validation by receipt; the validation receipts are the holdout used to score the Receipt Kitten adapters).
• synthetic_lines: generator (db, produce, meat, deli, bakery, nonfood, brand_only, fees, misc, fuel, menu, seed), upstream, name_v1/category_v1 (the generator's draft, taxonomy v1), relabelled (false = teacher did not answer; the draft is kept), split.
• household_synth: generator, chain (the chain whose POS style was imitated), item_id (same product -> same id), variant (0/1: two printed labels per product), generation_confidence, department_prompt.

How it was built

1. Receipts. Candidates from ReceiptBench, ExpressExpense SRD, JensWalter/my-receipts and docjay131/receipts-ocr-dataset (English purchase receipts only: stores, restaurants, fuel, parking). Near-duplicate images were removed with a 16x16 difference hash (<= 40 of 512 bits); ReceiptBench contains re-uploads of other datasets, so every ReceiptBench image that matches a WildReceipt image was dropped (WildReceipt has no stated license), and receipts matching a private evaluation set were removed by image hash and by shared item labels.
2. OCR. Apple Vision text recognition on the Mac; the item lines come from the source's own annotations where it has them (ReceiptBench) or from the OCR rows.
3. Teacher. DeepSeek V4.1 Flash (deepseek-flash, thinking effort low) read each receipt's lines with the store context and returned name + category + confidence per item; medium/low items got a second pass with thinking high. Only the item lines are published: payment, tender, loyalty and tax rows are not.
4. Synthetic lines. A generator abbreviates product names the way POS systems do (vowel dropping, store-brand codes, size tokens, truncation to the printer width of each chain). The drafts were re-labelled by the same teacher against taxonomy v2.
5. Household inventory. The teacher listed 40 products per chain and department and wrote two plausible printed labels for each. A second teacher pass re-read every generated label blind (store + label only); rows were kept only if it agreed on the category, was not low-confidence and did not rename the product.
6. Checks. Every row passed a regex scan for card numbers, phone numbers, e-mail addresses and staff/customer names; documents that carry personal names (phone bills, hotel folios, flight itineraries) were excluded as a class.

Quality

On a private hand-checked set of 988 line items from 152 real US receipts, the teacher's leaf category agrees with the hand labels 92% of the time (98.7% for items it marks high confidence). A Qwen3.5-4B LoRA trained on these data plus the ODbL part reaches 79.2% leaf / 90.1% top-level accuracy on that set (lxe/receipt-item-namer).

Limitations

• Labels are model-generated, not human-verified. Expect about 8% wrong leaves overall, more on medium/low rows and on boundary categories (candles, light bulbs, contact-lens solution, bread vs. bakery desserts).
• real_lines is about 80% restaurant and cafe; real US retail receipts with open licenses are scarce. Household and retail coverage comes mostly from synthetic data, whose printed labels are plausible rather than real.
• Names for heavily abbreviated labels are best guesses; brands are sometimes inferred from the store's own brand.
• US-centric (some CA/UK/AU/SG receipts in real_lines); prices in the source currency.

License and attribution

MIT (see LICENSE). Built from:

• ReceiptBench, T0rl/ReceiptBench on Hugging Face, MIT.
• ExpressExpense SRD receipt image dataset, MIT. Please cite ExpressExpense.com as the source (https://expressexpense.com/blog/free-receipt-images-ocr-machine-learning-dataset/).
• JensWalter/my-receipts (Kaggle jenswalter/receipts), CC0 1.0.
• docjay131/receipts-ocr-dataset on Hugging Face, MIT.
• USDA FoodData Central Branded Foods (2026-04-30 release), public domain (CC0 1.0).
• Teacher outputs from the DeepSeek Open Platform. DeepSeek's terms assign output rights to the user and permit training other models.

Upstream copyright notices are reproduced in LICENSE.