davidr99/us-address-standardization

US Address Standardization (PostGIS stdaddr schema)

Chat-format instruction data that maps a raw US address string to a strict JSON object matching the PostGIS address_standardizer stdaddr type (camelCase keys, USPS-abbreviated values). Used to fine-tune the qwen35-address-std model family.

Each row has three flat fields (system / user / assistant) for easy reading and grepping; rebuild the chat messages list from them at train time:

{
  "system": "<standardization instructions>",
  "user": "100 Old Forge Rd, Kent, CT 06757",
  "assistant": "{\"houseNum\":\"100\",\"qual\":\"OLD\",\"name\":\"FORGE\",\"suftype\":\"RD\",\"city\":\"KENT\",\"state\":\"CT\",\"country\":\"USA\",\"zipcode\":\"06757\"}"
}

Output contract (16 keys)

building, houseNum, predir, qual, pretype, name, suftype, sufdir, ruralRoute, extra, city, state, country, zipcode, box, unit

Empty fields are OMITTED (v5 onward): a missing key means "not present in the input", which trims output tokens ~43%. Consumers must treat a missing key as empty. sufdir is the post-directional (NE in "Main St NE"); extra carries floor info or an intersection's cross street (& 5TH AVE); ruralRoute is number-only; country is always USA.

Versioning

Pin a revision tag for reproducibility (load_dataset(repo, revision="v6")); main tracks the newest.

Tag	Rows (train / val)	What it adds
v5	48,896 / 2,139	omit-empty format; 100% US city coverage (~18.6k cities)
v6	50,966 / 2,369	production-audit failure families (all generated)
v7	58,742 / 3,233	full regen with the families in the base distribution; ZIP-transcription decoupling; realistic large unit numbers; state-name-city disambiguation
v8	59,606 / 3,329	numbered-PRIVATE-route rebalance
v9 (latest)	59,606 / 3,329	comma-delimited unit field; PRIVATE pretype constraint in the prompt; two v8 additions reverted

Earlier tags v1-v4 use the full-JSON format (every key always present).

v9: comma-delimited unit field

• Comma-delimited unit. The consuming service builds candidate strings as Street, Suite, City, State Zip -- the suite is its own comma-delimited field. That shape appeared once in all of v8 (0.002%) despite being the primary format the model is now sent; it is 6% of v9. The glued, comma-less form is deliberately retained (19% of rows) because external providers and consumer input still arrive without delimiters.
• PRIVATE is a pretype value, not a qualifier. A v8-trained model put the route number in name correctly but split the descriptor as qual=PRIVATE, pretype=RD -- a combination that appears zero times in the training data. It is composed from qual=OLD + pretype=AVE, both of which are legitimate patterns, and more volume did not displace it. The prompt now states the constraint explicitly.
• Two v8 additions reverted. Nine name-final type-words (TRAIL, WAY, ...) bought no measured improvement on the case that motivated them, and the St Rte State-Route surface caused 390 NORTHEAST ST 6 to parse as STATE RTE 6. Both were justified by one or two observations -- too thin for the capacity and risk they carried.

v8: numbered-PRIVATE-route rebalance

A model trained on v7 scored 96.5% exact-match on the v7 validation split but missed ~100% of numbered PRIVATE routes (94 pretype misses against exactly 94 PRIVATE rows), while the structurally identical RANCH RD family -- added at the same time -- scored fine. Two v7 data defects explained the difference, both fixed here:

• PRIVATE was in the ordinary street-name pool, generating rows that taught name=PRIVATE in direct competition with the route rule (a 4.5:1 route-to-name ratio). It is now emitted only as a deliberate, rare contrast case (31:1).
• PRIVATE routes could take a qualifier ("Upper Private Rd 2616"), which taught the model to read PRIVATE itself as the qualifier -- observed as "Private Road 834" -> qual=PRIVATE, pretype=RD. PRIVATE routes now take no qualifier.

The family's share of the dataset also doubles (1.47% -> 2.96%). "Private Drive" reads as a natural street name, so this rule has to overcome a strong prior and needs the extra support that RANCH RD did not.

v7 refinements (over v6)

v7 is a full regeneration (v6 appended the families to the older v5 base; v7 bakes them into the base generator and adds three data-quality fixes):

• ZIP is transcription, not recall. ~10% of non-coverage rows carry a random valid-format ZIP decoupled from the city's real ZIP, so the model learns to copy the input ZIP rather than recall the city's. Earlier splits could only contain gazetteer ZIPs, so their zipcode accuracy could not distinguish transcription from recall; the validation split now includes a genuine unseen-ZIP slice.
• Realistic unit numbers. Suite/apartment numbers now span 1-4 digits (Ste 402, Apt 1122), not just <= 40 -- the large-numeric-token-next-to-a-ZIP density behind number-dense parse slips.
• State-name-city disambiguation. A city whose name is a state (Nevada MO, Washington) is never emitted without its state token, removing a class that is inherently ambiguous with the bare-state reading.
• Numbered-route coverage. Descriptor families found by mining real US address data, each parsed as pretype + number-in-name (never number-in-unit): PRIVATE RD / PRIVATE DR ("Private Road 631" -- Holmes/Belmont county OH), RANCH RD (Texas), plus surfaces for Township Hwy, St Rte, SH, County Hwy/CTH, and bare Route N. Lettered county trunk routes ("County Highway FF", CTH KK) put the LETTER in name. A bare "Private Dr" with no number stays an ordinary street named PRIVATE. Upper-Midwest grid house numbers (N4451, N55W13775) are recognized.

Production-audit families (v6+)

The dataset targets failure modes found by replaying two production audits against the deployed model:

• ICOMS conversion audit (MCTV Block 3) - directional words inside multiword street names (SMITHVILLE WESTERN), state-name streets (OHIO STATE DR), direction-prefixed cities in the comma-less form (... Rd North Lawrence OH), suffix-less landscape names (WINCHESTER WOODS), fused dir/qual names (NOLD, SOUTHRIDGE), X AND Y names, Mc/Mac mixed case, city-pair rural roads, qualified/lettered routes.
• deepparse / libpostal US gaps - county phrases (Jefferson County, dropped from the output), alternate suffix abbreviations (Str/Crt/Drv/Lp/Wy/La), and unit designator variants (No/Nbr/Ap/Un/U, Door/Flat).

No real production address appears in train/val. The audit strings only motivate the synthetic families; every training row is rule-generated. The real addresses are held out as a separate real-world evaluation set (mctv_eval_holdout.jsonl) in the companion model repo.

Generation

Synthetic, label-first: a valid structured record is composed, then a deliberately messy raw string is rendered from it (varied casing, USPS abbreviation spellings, punctuation, partial city/state/zip tails). Covers streets, numbered routes, interstates, intersections, rural routes, PO boxes, named buildings, and tricky unit designators. A companion compare dataset (us-address-comparison) trains the same model to judge whether two addresses refer to the same place. Generator, coverage oracle, validity checker (validate_generated.py), and regression guard live in scripts/ in the companion model repo. For exact PostGIS fidelity, labels can instead be produced with postgis_label.py against a live PostGIS instance.