This model is finetuned and quantized based on a pretrained huggingface BERT model. The quantization method is: per-tensor, symmetric, zero_point=0. It uses NVIDIA's quantization toolkit on top of PyTorch to perform quantization. Achieved accuracy is f1_score=90.633%. A description of the quantization steps can be found in README.md. All code necessary to reproduce can be found in the upload: Dockerfile, run_squad.py, quant_trainer.py, and modeling_bert.patch. The PyTorch model itself is pytorch_model.bin.