BERT-based-MRC

This model is a fine-tuned version of bert-base-chinese on BD-11 dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 18
• optimizer: Adam with epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3

Training results

Evaluated on the BD-11 dataset test set.

  "exact": 87.70,
  "f1": 89.32,
  "total": 1528,
  "HasAns_exact": 83.94,
  "HasAns_f1": 86.24,
  "HasAns_total": 1077,
  "NoAns_exact": 96.67,
  "NoAns_f1": 96.67,
  "NoAns_total": 451

Framework versions

• Transformers 4.25.1
• Pytorch 1.6.0
• Datasets 2.7.1
• Tokenizers 0.11.6