Fuzzy Matching Test Data: A Downloadable Dataset With Known Matches

What fuzzy matching is — and isn't

Fuzzy matching is one technique used within entity resolution, not a synonym for it. It answers "are these two strings similar enough to be the same thing?" — entity resolution is the broader question of "do these records refer to the same real-world entity?", which can combine fuzzy matching with blocking/indexing and deterministic rules. For the full-record version of that broader question, see the duplicate customer records dataset.

Algorithm glossary

Levenshtein distance

The minimum number of single-character edits (insertions, deletions, substitutions) to turn one string into another.

Jaro / Jaro-Winkler

Similarity scores tuned for short strings like names, giving extra weight to matching prefixes (Jaro-Winkler).

Soundex / phonetic matching

Encodes words by how they sound rather than how they're spelled, catching phonetically similar but differently spelled names.

Edit distance

General term for the number of operations needed to transform one string into another (Levenshtein is one specific edit-distance metric).

N-grams

Overlapping substrings of length N, compared as sets — robust to reordering and partial matches.

Token-based matching

Splits a string into tokens (words) and compares the token sets, so word order doesn't affect the match.

Precision / recall / similarity threshold

The tuning tradeoff: a looser threshold catches more true matches (recall) at the cost of more false positives (precision).

Sample dataset

Representative rows from the full 250-row file (50 identity groups):

Data dictionary

Example test cases

Expected-match answer key

Methodology

Generated with Generate-Data's schema builder in duplicate-generation mode: one master record per identity, then variant records tagged against it via Master ID / Duplicate Type columns. This is synthetic data — no real people, no real PII — and it's not a benchmark of any specific algorithm's accuracy; it's a labeled fixture for testing your own matching logic.

Generate a larger version

Need more rows, more identity groups, or a different variant mix? Open the schema builder → add Name/Email fields → switch on duplicate generation mode → set your duplicate rate and Duplicate Type mix → export as CSV, JSON, XML, or Parquet (JSONL/ HuggingFace also supported). AI-assisted field suggestions, if used to help design field values, are capped at 1,000 rows per batch — the duplicate-generation engine itself isn't subject to that cap.