sraivante/zip-county-city-data

ZIP County City data - v0.3.0

Actual saved training and evaluation data for model v0.3.0, run internal_v3_20261021_01, completed on L4 on 2026-10-21. The paired model is 31.49M parameters, trained from random weights, with no separate fine-tuning data. The same 5,000-prompt complete-response score increased 25.74% -> 68.96%, while fresh-wording complete output is 12.10% and intent regressed. See the model card and evaluation/RESULTS.md for complete limitations.

Snapshot and contents

Artifact	Saved rows	Role
parser train	360,000	1,536,000 presentations across all 6,000 updates
parser validation (original dev)	4,000	checkpoint selection every 500 steps
parser test	10,000	locked before training; final wording evaluation
canonical geography.npz	41,089	all known ZIP facts; 6,144,000 presentations
membership_county.npz	254,178	6,144,000 presentations
membership_city.npz	334,926	6,144,000 presentations
membership_state.npz	205,445	6,144,000 presentations
legacy regression	5,000	diagnostic evaluation only
expanded regression	59,952	diagnostic evaluation only
source geography CSV/Parquet	98,597	41,089 ZIPs, 59 state/territory labels

archive/training_input.zip is the exact input uploaded for the successful L4 run, including original JSONL/NPZ arrays, code, baseline v2 evaluation model and historical cards. Extract it for byte-identical paths and evaluation reproduction. Separate gzip JSONL files decompress to the original bytes; compression_manifest.json records original hashes and counts. NPZ, vocab, templates and the original data manifest are separately available under data/model_v3/. All four saved prediction files (v3 legacy/expanded/fresh and v2 fresh) are under evaluation/. No source data was regenerated or edited after training. Later analysis sidecars are evaluations, not changes to the training snapshot. The original optimizer/RNG resume state and private account/Drive files are excluded from this public release.

from datasets import load_dataset
ds = load_dataset("sraivante/zip-county-city-data", "parser", revision="v0.3.0")
print(ds["train"][0])

ZIP columns are strings, preserving leading zeros. geography reads typed Parquet; the byte-identical CSV is also included. Membership arrays can be loaded with numpy.load(..., allow_pickle=False)["rows"] after download.

Schema and generation

Parser JSONL fields: id, English text, five-digit zip, given (optional county/city/state strings), spans (ZIP/county/city/state order), task, template, expected (state/county/city), and boolean valid. Spans are inclusive start/end positions after an initial null token at index zero; missing fields use [0,0]. NPZ contains input character IDs, spans, integer ZIP, name IDs, field presence, intent and validity targets. Output names are uppercased. Character inputs are lowercased and limited to 160 characters. Raw copied names are not spelling fixes. Geography NPZ stores one canonical row per ZIP. Membership NPZ rows contain ZIP index, name index, binary compatibility. Vocabularies map bare names to IDs.

Seed 20261021; lookup, pair, state, county-only and city-only families each have 72,000 training contexts. Positive/negative examples share constructions. All accepted names per ZIP are aliases; representatives use the last primary row, else first. County/city negatives are outside all aliases in the true state; state negatives use another state. Membership targets include every positive alias, three sampled negatives per positive where available, plus an explicit unknown-name negative per ZIP. A separate learned validity head is trained on seeded synthetic confidence/truth/presence patterns generated by the versioned trainer, not materialized geographic rows.

Whole template strings and normalized texts have zero overlap across parser train/development/test. All ZIP facts are seen in training; this is a wording generalization split, not an unseen-geography split. Legacy/expanded test sources and hashes are recorded in data/model_v3/manifest.json; their original contexts and labels are retained. Old benchmarks informed diagnosis. The fresh test was not used for checkpoint selection; selected step was 5,500 (24.925% development complete response). Post-test gold-parser analysis is diagnostic and not an achieved end-to-end score. Future tuning needs a new locked test.

Identity, history, license and provenance

Data manifest SHA-256: f3b89fdc90fd2133e443991315413d3ce25c3a24670be8cfd4b88bfc3f82137c. Training input SHA-256: b18ce0f286808410bb40fe66821685be605e9c01ae704257a7ce42fedbb14f5b. Source SHA-256: 9ca02a431e2af1429f9cf64942ef741a3dd5223a8e8b1e635ed7c371c86ccf9b. Paired model SHA-256: 8b36d48fd06f34048238756cfe98dc1e7248dd0c71452119916eb57acaf33e5e. provenance.json and release_manifest.json record exact counts, file/code hashes, baseline revisions and uncertainty. No Git source commit was recorded. Archive cards retain preparation-time private/untrained status for historical identity; the current cards record completed training and this authorized public release.

Source: owner's MySQL state_county_city_zipcode export, identical to v0.2.0. Upstream provider, collection date, revision and license are unrecorded. No third-party notices were supplied; no nonredistributable source is identified in available documentation. This gap is disclosed rather than asserting verified Apache ownership of upstream facts or current authoritative postal status.

Copyright (c) 2026 sraivante for original code, templates, annotations and original selection/arrangement, under Apache License 2.0. Preserve third-party ownership and terms; see NOTICE. No base-model license applies. Source names may have ambiguous or stale representatives, and learned results remain unreliable for unfamiliar wording, arbitrary spelling, unsupported or out-of-scope inputs.

Previous datasets remain at v0.1.0 and v0.2.0.