Downloads · 30 days
88
3% of all-time downloads
madlag/bert-large-uncased-whole-word-masking-finetuned-squadv2
bert-large-uncased-whole-word-masking-finetuned-squadv2 is a question answering model from madlag. Use it when the input is a question plus a passage. It is set up for transformers.
Used run.sh used to train using transformers/example/questionanswering code.
Downloads · 30 days
88
3% of all-time downloads
All-time downloads
2.9K
Public
Repo size
4 GB
Likes
0
Public
Click a slice to open those files.
.bin1.3 GB · 50%
From the Hugging Face model README
Used run.sh used to train using transformers/example/question_answering code.
Evaluation results : F1= 85.85 , a much better result than the original 81.9 from the BERT paper, due to the use of the "whole-word-masking" variation.
{
"HasAns_exact": 80.58367071524967,
"HasAns_f1": 86.64594807945029,
"HasAns_total": 5928,
"NoAns_exact": 85.06307821698907,
"NoAns_f1": 85.06307821698907,
"NoAns_total": 5945,
"best_exact": 82.82658131895899,
"best_exact_thresh": 0.0,
"best_f1": 85.85337995578023,
"best_f1_thresh": 0.0,
"epoch": 2.0,
"eval_samples": 12134,
"exact": 82.82658131895899,
"f1": 85.85337995578037,
"total": 11873
}