--- license: mit language: - en pretty_name: HomeGoods Taxonomy size_categories: - n<1K task_categories: - text-classification - zero-shot-classification tags: - taxonomy - classification - inventory - logistics - moving - household - dangerous-goods - hazmat configs: - config_name: default data_files: - split: train path: taxonomy.csv --- # HomeGoods Taxonomy v1 A hierarchical classification of the physical goods in an ordinary home or small business, with a **physical handling attribute layer** attached to every class. Modelled on the Harmonised System — HS codes classify goods for customs, and answer *what duty is payable*. This answers a different question: > **What does this object need from the box it is in, and what should it be > stored next to?** ``` 30 Electronics & Computing 3008 Storage Media 300801 Internal drives 0.5 kg · 0.65 kg/L · fragile · high value · nothing on top ``` **367 classes** across **136 groups** and **27 divisions**. Every item in an inventory gets one 6-digit code, and the attributes come with it. ## Why An inventory row records *what you have*. It does not record that a box of hardbacks will weigh 24 kg, that bleach must never be packed with drain cleaner, or that the button cells in a drawer are the most dangerous thing in the house to a toddler. Those are properties of the **kind** of object, not of the object — so they only need recording once. Record them per class, assign classes automatically, and an existing inventory acquires a data layer it never had. ## Files | File | Shape | |---|---| | `taxonomy.csv` | Flat, one row per class, 36 columns. The default config. | | `taxonomy.json` | Nested — divisions → groups → classes — plus the defaults map | Both are generated from the same YAML source. ## Loading ```python from datasets import load_dataset ds = load_dataset("danielrosehill/homegoods-taxonomy", split="train") # Everything a removal firm will refuse refused = ds.filter(lambda r: r["mover_restricted"]) # Everything that has to be hand-carried hand_carry = ds.filter(lambda r: "hand_carry" in (r["handling"] or "")) ``` `aliases`, `child_hazard_reasons`, `co_store_avoid` and `handling` are pipe-delimited (`a|b|c`) in the CSV and real arrays in the JSON. ## The code scheme Fixed-width numeric, where code length is the level and every code is a prefix-extension of its parent. | Level | Digits | Example | |---|---|---| | Division | 2 | `30` Electronics & Computing | | Group | 4 | `3008` Storage Media | | Class | 6 | `300801` Internal drives | `LIKE '30%'` selects an entire division — no recursive join needed to roll a box up. **Codes are strings, not integers**; store as `TEXT`. Codes are never reused and never renumbered. `999901` is the code a classifier emits when it is **not confident**. A populated bucket of those is a work queue; a wrong code is a lie that propagates into every roll-up built on it. ## Attributes Divisions and groups are structure. Only classes carry the payload. **Physical** — `typical_mass_kg`, `bulk_density_kg_per_l`, `nestable`, `stackable`, `load_bearing`, `orientation` **Sensitivity** — `fragile`, `crushable`, `moisture_sensitive`, `temperature_sensitive`, `light_sensitive` **Safety** — `child_hazard`, `child_hazard_reasons`, `hazmat`, `un_class`, `un_indicative`, `battery`, `perishable` **Logistics** — `storage_zone`, `access_frequency`, `two_person`, `disassembly`, `handling`, `value_density`, `mover_restricted`, `co_store_avoid` ### Bulk density is the field that does the work Item mass tells you what one thing weighs. **Bulk density tells you whether the box is liftable**, which is the constraint that actually bites. A 40 L box has a safe manual-handling ceiling around 20 kg, so **0.5 kg/L is the line**: | Class of thing | ~kg/L | A 40 L box weighs | |---|---|---| | Insulation | 0.03 | 1 kg | | Duvets, pillows | 0.05 | 2 kg | | Folded clothing | 0.15 | 6 kg | | Cables, coiled | 0.35 | 14 kg | | Paperbacks | 0.50 | 20 kg ← the limit | | Hand tools | 0.65 | 26 kg | | Paper, documents | 0.75 | 30 kg | | Tinned food | 0.90 | 36 kg | | Screws & fixings, loose | 1.20 | 48 kg | | Weight plates | 2.50 | 100 kg | The classic packing mistake — a large carton filled with books — is one multiplication away from being caught. ### `co_store_avoid` is the only field about combinations Solvents do not go in with textiles; nothing corrosive near electronics; nothing chemical shares a box with food. This field exists because similarity engines get it actively wrong. An affinity model that clusters on similarity will happily put drain cleaner next to the flour — both live in the kitchen, both are used monthly, both weigh about a kilogram. This is what stops it. ### `handling` is actions, not fragility `fragile` says the thing breaks and calls for padding on the day. `handling` says something must be **done** to it, or **by someone specific**, and most of it has a lead time: `hand_carry` · `drain_required` · `depressurise_required` · `transit_restraint` · `settle_after_transit` · `professional_disconnect` · `specialist_mover` · `rigid_case` · `climate_controlled` · `customs_declaration` · `permit_required` · `living` · `magnetic` A washing machine moved without its transit bolts destroys its own drum bearings. A pet's rabies waiting period runs to months. Neither is a packing problem. ### Dangerous goods `hazmat` is plain language for the person packing the box. `un_class` is the UN/ADR hazard class for the mover or freight forwarder. Both are set, or neither is — the pair is validated. 38 classes are `mover_restricted`. 54 carry a UN class, 22 of them class 9, which is almost entirely lithium batteries. > **This is not a shipping declaration.** A real declaration needs the specific > UN number, proper shipping name, packing group and quantity, determined from > the actual product and its safety data sheet. What this does is surface, weeks > ahead, which items a remover will refuse — so they can be used up, given away > or disposed of on a schedule rather than discovered on the morning. ## Intended use Built to be assigned by a model, not by hand. Nobody classifies a thousand items manually, so class descriptions are written to be read by a classifier as much as by a person — they say what falls *outside* a class wherever the boundary is not obvious. Good fits: - **Inventory enrichment** — assign a code per item, inherit the attributes - **Box and container roll-ups** — "this box will weigh 26 kg", "this box holds a hazmat item", "this shelf is within a child's reach and should not be" - **Move planning** — weight limits, refused-goods lists, hand-carry sets, unpacking order, lead-time tasks - **Storage layout** — cluster by zone, access frequency and attribute profile, subject to the anti-affinity constraints Poor fits: retail categorisation, customs classification, insurance valuation, anything needing per-item accuracy rather than per-class planning figures. ## Depth Set for organisation and moving, not for completeness. A class exists where it changes how something is stored, packed or carried — which is why "gardening gloves" and "smart plugs" are classes and the difference between two brands of screwdriver is not. Stopping there is deliberate. A decorative taxonomy of everything in a home is unbuildable, and past a certain depth it stops changing any decision. Level 4 (8-digit subclasses) is reserved for consumers who need to go further in one branch. ## Caveats **All mass and density figures are estimates**, accurate to roughly ±30% — enough to keep a box under a weight limit, not enough to bill a shipper on. Where a class spans a wide range, the figure is the middle of the common domestic range, not the mean of everything sold. Attributes do **not** inherit down the tree. A group is not required to be homogeneous, and every class states its own values. Storage zones and packing conventions reflect a European/UK domestic norm. Some of it will not transfer — cupboard names, which appliances are fitted, what a removal firm's standard terms exclude. ## Source Developed at [github.com/danielrosehill/HomeGoods-Taxonomy](https://github.com/danielrosehill/HomeGoods-Taxonomy), which holds the YAML source of truth, the validator, the build tooling, and the reasoning behind the schema. This dataset is the published release. ## Licence MIT.