| Model | fused | quantized | pruned | Score (F1) | Time (s) |
|---|---|---|---|---|---|
| PyTorch RuBERT | No | No | 0% | 67.26% | 316 |
| ONNX RuBERT | No | Yes | 0% | 67.25% | 368 |
| ONNX RuBERT | No | Yes | 50% | 59.17% | 332 |
| PyTorch Truncated LaBSE (en-ru) | No | No | 0% | 70.01% | 317 |
| ONNX Truncated LaBSE (en-ru) | No | Yes | 0% | 13.91% | 374 |
Not every huggingface model can be quantized efficiently, but for some models quantization can be done without any loss in modelling capabilities.