roberta-base-squad2-nq-bioasq

Model description

This model is a fine-tuned version of nlpconnect/roberta-base-squad2-nq on the BioASQ 10b dataset.

Intended uses & limitations

Cross-domain question answering!

Training and evaluation data

Training: BioASQ 10B with SQUAD sampled evenly to match the same samples as BioASQ 10B Eval: BioASQ 9B Eval with SQUAD Eval sampled evenly to match the same samples as BioASQ 9B Eval

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Went from untrained exact match: 60.9% (f1 71.8%) to exact match: 95.2% (96.6% f1) on BioASQ 9B held out training set. Scores on SQUAD+BioASQ remained stable at exact match: 72.5% (f1 81.4%) to 88.5% (f1 93.3%).

Framework versions

• Transformers 4.18.0
• Pytorch 1.11.0
• Datasets 2.1.0
• Tokenizers 0.12.1