danielrosehill/inventory-spec-lines

Inventory Spec Lines

14,426 product-specification lines from a real household inventory (cables, adapters, tools, electronics, household goods), each labelled with spans for a fixed set of 25 attributes (vocabulary v2): manufacturer, model_number, mpn, serial_number, version, power_source, input_voltage, input_current, output_voltage, output_current, rated_voltage, rated_current, power, battery_capacity, battery_type, connector, plug_type, cable_length, diameter, volume, storage_capacity, pack_count, material, colour, ip_rating.

serial_number is in the vocabulary but has no examples here on purpose: every serial-keyed line is removed as PII. The model learns it from generated serials at training time (train/synthetic.py in the code repo).

The dataset was built to train a small token-classification model that turns free-text specs ("Output: 5V DC, 1A") into structured values. The model is in training and will be published as danielrosehill/inventory-spec-extractor, with its code.

Format

One JSON object per line:

{"id": "3f9c0a1b2c3d4e5f:2", "item": "3f9c0a1b2c3d4e5f", "product_type": "Power Adapter",
 "line": "Output: 5V DC, 1A",
 "spans": [{"start": 8, "end": 13, "text": "5V DC", "attribute": "output_voltage"},
           {"start": 15, "end": 17, "text": "1A", "attribute": "output_current"}]}

start/end are character offsets into line. item groups the lines of one inventory item; it is an opaque hash. A line with nothing to label has spans: []. That is deliberate: the model needs negatives. vocab.json defines each attribute, with its canonical unit and, for enums, the allowed values.

Splits are by item (no item appears in both): train 12,949, validation 1,477. 4,647 lines carry at least one span.

attribute	spans
material	1987
connector	1224
colour	1012
power_source	321
plug_type	180
battery_type	143
manufacturer	124
rated_voltage	107
pack_count	107
diameter	105
input_voltage	94
output_voltage	92
power	89
volume	84
storage_capacity	80
output_current	70
rated_current	64
model_number	63
cable_length	62
version	38
battery_capacity	30
ip_rating	24
input_current	9
mpn	6

How it was made

• Lines: the specs field of a personal inventory, split one line per row. Most specs were drafted by an LLM from item photos and titles during cataloguing, so the text is product facts, not personal writing.
• Labels: anthropic/claude-haiku-5.5 (OpenRouter batch, 2026-10-15), prompted with the attribute list and explicit span-boundary rules. The model returned span text; offsets were computed by exact substring match, and spans that did not match were dropped.
• These labels are not human-verified. In a 200-line pilot, Claude Haiku 5.5 agreed best with a 3-model majority (precision 0.958, recall 0.932). A hand-checked gold set is planned but does not exist yet. Treat the labels as silver.
• Span convention: the shortest text naming the value. Qualifiers are left out ("HDMI (Female)" → HDMI, "3-pin XLR Male" → XLR), and a list gives one span per value ("Silicone/Rubber" → two material spans).

Privacy

Every line was scanned before release (on a Modal T4) with Microsoft Presidio (e-mail, phone, IP, card and IBAN patterns), the GLiNER PII model urchade/gliner_multi_pii-v1 (people's names and street addresses), and inventory-specific rules (MAC addresses, IMEIs, and any line keyed serial / S/N / licence key / password / SSID). A second pass waived known false positives (IP ratings read as IP addresses, connector genders read as names) and added a private list of the household's own names. Flagged lines were removed, not masked: 68 lines. Only the spec line and product type are published: no item names, notes, serials, locations, prices or internal identifiers.

Known limits

• Electrical attributes are sparse (tens to about 100 spans each), because most household items have no voltage. input_current has fewer than 10.
• English only. Hebrew printed text was transcribed into English during cataloguing.
• Silver labels: see above.

Built 2026-10-15.