Name Matching on 50,000 real-world pairs drawn from historical production data (held-out set)
99.02AccuracySGER
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SGERBackbone=Llama 3 8B Instruct2026.05 | 99.02 | 99.95 | 98.9 | 99.4 | |
| LLMEvaluation Protocol=SFT, Backbone=Llama 3 8B, Data Augmentation=true2026.05 | 95.7 | 97.6 | 97 | 97.3 | |
| LLMEvaluation Protocol=SFT, Backbone=Llama 3 8B2026.05 | 91.2 | 93.3 | 95.9 | 94.6 | |
| GPT-4oEvaluation Protocol=Few-Shot2026.05 | 85.4 | 90.1 | 92.1 | 91.1 | |
| LLMEvaluation Protocol=Few-Shot, Backbone=Llama 3 8B2026.05 | 74.3 | 85.2 | 82.6 | 83.9 | |
| BERTEvaluation Protocol=Fine-Tuned, Backbone=bert-base-multilingual-cased2026.05 | 69.1 | 81.2 | 80.1 | 80.6 | |
| LevenshteinThreshold=0.82026.05 | 57.4 | 75.1 | 70.2 | 72.6 |