IPC Classification on Patent dataset IPC
48.8Macro F1TechToken BERT
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| TechToken BERTInference approach=next-token probability thresholding, Selection Strategy=Best across mean-pool and CLS embedding2026.05 | 48.8 | 44.9 | 0.02 | |
| PaecterSelection Strategy=Best across mean-pool and CLS embedding, Inference approach=k-nearest neighbor [36]2026.05 | 42 | 37.8 | 0.1 | |
| PatentSBERTaSelection Strategy=Best across mean-pool and CLS embedding, Inference approach=k-nearest neighbor [36]2026.05 | 35.6 | 30.3 | 0.12 | |
| BERT4PatentsSelection Strategy=Best across mean-pool and CLS embedding, Inference approach=k-nearest neighbor [36]2026.05 | 35.4 | 30.1 | 0.12 | |
| LLaMA PatentsSelection Strategy=Best across mean-pool and CLS embedding, Inference approach=k-nearest neighbor [36]2026.05 | 34.3 | 29.5 | 0.12 | |
| BERT4Patents FTSelection Strategy=Best across mean-pool and CLS embedding, Inference approach=k-nearest neighbor [36]2026.05 | 26.2 | 22.5 | 0.15 |