inventory-spec-extractor

A small model (33M parameters, int8 ONNX, about 34 MB) that finds structured attributes in one line of a free-text product specification:

input line	output spans
Output: 12V 3A	12V output_voltage, 3A output_current
Length: 2m, USB-C to USB-A	2m cable_length, USB-C connector, USB-A connector
Inlet: IEC C14, Outlets: 8x C13	IEC C14 connector, C13 connector
Plug: Israeli (Type H), Power: 2200W	Type H plug_type, 2200W power
Power Source: 2x AAA batteries	AAA battery_type, batteries power_source
S/N: C02XG1ABJG5H	C02XG1ABJG5H serial_number

The model finds and labels spans; it does not convert units or pick enum spellings. Do that in deterministic code: "100-240V AC" → 100–240 V, AC. The model is built to run on CPU in the browser or on a small server through onnxruntime-web, at about 20 ms per line.

Labels (25 attributes, BIO)

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

Input text is {product_type} | {line} (product type may be empty). Labels sit on the first word-piece of each word; continuation pieces extend the open span. labels.json maps logit index → O / B-attr / I-attr.

Evaluation

Validation split of inventory-spec-lines (1477 lines, split by item, never trained on). Scoring is span-level exact match on (start, end, attribute). The eval labels are silver (Claude Haiku 5.5, not hand-checked), so these numbers measure agreement with that teacher. The baseline is a regex extractor (key names plus number+unit patterns).

	precision	recall	F1
regex baseline	0.804	0.454	0.580
this model	0.834	0.860	0.847

attribute	n	baseline F1	model F1	model P	model R
material	184	0.54	0.93	0.93	0.92
connector	140	0.36	0.84	0.78	0.90
colour	72	0.82	0.88	0.82	0.94
battery_type	26	0.27	0.77	0.86	0.69
output_voltage	26	0.67	0.83	0.95	0.73
power_source	26	0.00	0.78	0.80	0.77
plug_type	24	0.70	0.77	0.78	0.75
output_current	20	0.97	0.63	0.67	0.60
input_voltage	15	0.73	0.77	0.75	0.80
rated_voltage	15	0.80	0.97	0.94	1.00
storage_capacity	12	0.79	0.72	0.69	0.75
manufacturer	11	0.00	0.70	0.67	0.73
cable_length	10	0.63	0.91	0.83	1.00
volume	10	0.76	1.00	1.00	1.00
pack_count	9	0.18	0.78	0.64	1.00
power	8	0.93	0.80	0.86	0.75
diameter	6	0.80	0.77	0.71	0.83
version	6	0.00	0.29	1.00	0.17
battery_capacity	5	1.00	1.00	1.00	1.00
rated_current	4	0.80	0.67	0.60	0.75
input_current	3	1.00	0.00	0.00	0.00
model_number	2	0.00	1.00	1.00	1.00
ip_rating	1	1.00	1.00	1.00	1.00

Known weaknesses

• Currents. output_current, rated_current and input_current score below the regex baseline. On mixed lines ("Power: 12V, 3A, 36W") the model can label an amp value as power. A unit check in the caller catches this: an amp value is never a power.
• Decimals in some contexts. "0.5A" has been seen cut to "0".
• version is weak (few examples), and "Gen 2" can come out as model_number.
• mpn has only 6 training examples.
• serial_number is learned only from generated serials, because real serial lines are removed from the public data. It handles Serial Number: / S/N: / SN: lines well; it has not been measured on real serials.

Training

Fine-tuned from BAAI/bge-small-en-v1.5 for 6 epochs on a Modal T4. Lines with attributes that have fewer than 300 spans were repeated 4×, and 300 synthetic serial lines were added. Trained 2026-10-15. Code, labelling and redaction pipeline: danielrosehill/Inventory-Spec-Extractor.

Files

model_quantized.onnx (int8, what to load), tokenizer.json, labels.json. Inputs: input_ids, attention_mask, token_type_ids (int64 [batch, seq]). Output: logits float32 [batch, seq, 51].