gagan1985/indian-addresses-gold

Indian Addresses (Gold Standard)

Span-labeled Indian addresses — the training data behind gagan1985/qwen3-0.6b-indian-address-parser. For the raw, unlabeled corpus, see gagan1985/indian-addresses-raw.

4,834 records, split by provenance:

reviewer	Count	Description
llm:deepseek/deepseek-v4-pro	4,825	LLM-reviewed via OpenRouter, span-verified against the raw address text (a field value not found verbatim in the source string is dropped, not guessed)
gagan	9	Human-reviewed via Label Studio

source_type	Count
mca	2,508
bank	2,326

Fields

13 structured fields, one column each: houseNumber, houseName, poi, street, subsubLocality, subLocality, locality, village, subDistrict, district, city, state, pincode (null/None when not present in the address).

Plus:

Field	Description
record_id	Source-specific identifier
source_type	mca or bank
raw_address	The raw address string these fields were extracted from
review_state	REVIEWED_CONFIRMED (reviewer's spans matched the rule-based prediction exactly) or REVIEWED_CORRECTED (reviewer changed at least one span)
reviewer	gagan (human) or llm:<model> (LLM-reviewed) — always check this if you care about provenance
reviewed_at	ISO timestamp
rule_confidence_at_review_time	The rule-based pipeline's own confidence score for this record before review
spans	The canonical span-offset format: a list of {"start": int, "end": int, "label": str} dicts, character offsets into raw_address. The per-field columns above are derived from this — spans is the source of truth if the two ever appear to disagree (e.g. a field with multiple spans is joined with a space in the flat column).

A note on label quality

99.8% of these labels (4,825 / 4,834) come from an LLM review pass, not human review — see the reviewer field to filter to human-only if that matters for your use case. The LLM's own accuracy was spot-checked against a small (9-example) human-reviewed sample before scaling up, and known systematic gaps are documented in the model card's "known limitations" section (locality/subLocality/subsubLocality/village field-boundary ambiguity, street sometimes over-absorbing multi-part location clusters).

PII handling

174 bank-source records that contained embedded customer PII (phone numbers, S/O/D/O/W/O/C/O relational name markers referring to a private individual) were dropped rather than redacted in place — the spans offsets are character positions into raw_address, and redacting text changes string length, which would silently shift every downstream offset (some of which, e.g. poi, can legitimately overlap the affected region). Dropping was the safer choice for a curated gold set; see the raw dataset for the redaction-based approach used there instead.

MCA records containing C/O <name> were not dropped — there it names a company director, already public via MCA's own CIN disclosure, and is often itself the correct poi label.

Related

• Raw corpus — 4.37M unlabeled addresses this gold set was drawn from
• Model — Qwen3-0.6B LoRA fine-tuned on this data
• Python package — pip install indian-address-parser
• GitHub (includes a benchmark comparison against another Indian address NER model)