Project · 2026 · interactive
Is this the same business?
Three sources describe millions of businesses with no shared IDs: names abbreviated, transliterated or reordered, addresses partial. Some records are traps, a real business copied with one detail changed. The task was to find every record of each business while being penalised more for false matches than for missed ones. Below, the pipeline's first stages run on records you can edit.
- Public leaderboard rank, of 89,000+
- 696
- Score, macro F0.5
- 0.9839
- Records to match
- 10M+
- Countries, one with no training data
- 3
Try it
Reference business
Candidate record
Loading the word map…
The full pipeline: click a stage
LightGBM over 39 features like the ones above, plus five fine-tuned cross-encoders reading both records as text, combined by a second LightGBM.
Pick an example, decide for yourself, then look at what the pipeline sees. Edit any field: change a house number by three, or Inc to LLC, and watch the decoy signs light up. The records are made up; no challenge data is on this page.
Results
| Metric | Value | Source |
|---|---|---|
| Public leaderboard: rank / overall score (macro F0.5) | 696 / 0.983926 | README, leaderboard |
| Retrieval: true matches among the candidate pairs (train) | 98.0% | README |
| Pruner: candidate pairs kept on test | 125M → 11.6M | README |
| Cost of pruning, F0.5 | 0.0005 | README |
Team project. Details, including what did not work, are in the repository.