Abt-Buy
Harbor adapterThe Abt-Buy dataset for entity resolution derives from the online retailers Abt.com and Buy.com.
Auto71%Coding
- Samples
- 1,916
- Released
- 2010
- Scores
- 7
Billiger.de Products
Billiger.de Products: A Bilingual Entity Matching Benchmark
A bilingual German and English entity matching benchmark for e-commerce covering 13 product categories with fixed splits along corner cases, unseen entities, and development set sizes.
How to run
ltb · billiger-de-products
$ npm i -g lettokenburn-cli$ ltb bench info billiger-de-products
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
9 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.2 Rule-Guided | 81.6 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| GPT-5.2 Simple | 70.4 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| XLM-R | 63.7 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Ditto | 60.4 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| HierGAT | 58.4 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| RoBERTa | 57.9 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| R-SupCon | 47.0 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| WordCooc | 43.6 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
| Magellan | 35.7 | N/A | Independent | medium | Sep 29, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
Content & Construction: The benchmark consists of offers from the German price comparison platform billiger.de. It includes a German version (13,568 offers) and an English translation where pairs/splits are kept fixed. Data is constructed with three dimensions: corner-case ratios (20%, 50%, 80%), unseen entity fractions (0%, 50%, 100%), and development set sizes (small, medium, large), yielding 27 variants.
Scoring & Findings: The benchmark was validated using six supervised matchers and zero-shot GPT-5.2. Results indicate the dataset is more difficult than existing English benchmarks like WDC Products. Most PLM-based matchers perform better on English data; under German-only training, they lose performance on cross-language pairs compared to same-language pairs.