README.md

---
license: mit
language:
- en
task_categories:
- text-generation
tags:
- address
- address-comparison
- address-matching
- record-linkage
- us-addresses
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: validation
    path: validation.jsonl
---

# US Address Comparison (per-part match verdicts)

Chat-format instruction data for judging whether **two** US addresses refer to the
same place, part by part. It is the companion to
[`davidr99/us-address-standardization`](https://huggingface.co/datasets/davidr99/us-address-standardization):
each address is provided as the raw string **plus** its standardized JSON (the
output of the normalization model).

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

```json
{
  "system": "<comparison instructions>",
  "user": "Address A: ...\nA components: {...}\nAddress B: ...\nB components: {...}",
  "assistant": "{\"overall_estimate\":{\"match_type\":\"MATCH\"}, ...}"
}
```

## Output (7 buckets, in order)

`overall_estimate, street, house_number, city, state, zip, unit`

Each is `{"match_type": "MATCH" | "PLAUSIBLE" | "NOT_MATCHED"}`.

- **MATCH** - equivalent after standardization (two empty values also match).
- **PLAUSIBLE** - cannot be ruled out: one side missing the field, a minor spelling
  variant, a ZIP+4 vs the matching 5-digit ZIP, a house number differing only by a
  letter/"1/2", or a street matching by name but differing in type/directional.
- **NOT_MATCHED** - both present and clearly conflicting.

Field folding: `building`/`extra` -> street; `box`/`ruralRoute` -> house_number;
`country` ignored. `overall_estimate` is critical-field weighted: NOT_MATCHED if any
of street/house_number/city/state/zip is NOT_MATCHED, else MATCH only if every bucket
matches, else PLAUSIBLE (a unit-only difference never forces NOT_MATCHED).

## Generation

Synthetic, label-first: a complete address A is built, address B is derived
(identical / independent / controlled per-field mutations), and the pair is labeled
by a deterministic comparator (`compare_schema.compare_records`) that encodes the
rules above. Address A is always complete; only B may be missing fields. See the
`scripts/` folder in the companion model repo for the generator.