Overview

Language model: deepset/roberta-base-squad2-distilled
Language: English
Training data: SQuAD 2.0 training set
Eval data: SQuAD 2.0 dev set
Infrastructure: 1x V100 GPU
Published: Apr 21st, 2021

Details

• haystack's distillation feature was used for training. deepset/bert-large-uncased-whole-word-masking-squad2 was used as the teacher model.

Hyperparameters

batch_size = 6
n_epochs = 2
max_seq_len = 384
learning_rate = 3e-5
lr_schedule = LinearWarmup
embeds_dropout_prob = 0.1
temperature = 5
distillation_loss_weight = 1

Performance

"exact": 68.6431398972458
"f1": 72.7637083790805

Authors

• Timo Möller: timo.moeller [at] deepset.ai
• Julian Risch: julian.risch [at] deepset.ai
• Malte Pietsch: malte.pietsch [at] deepset.ai
• Michel Bartels: michel.bartels [at] deepset.ai

About us

We bring NLP to the industry via open source!
Our focus: Industry specific language models & large scale QA systems.

Some of our work:

• German BERT (aka "bert-base-german-cased")
• GermanQuAD and GermanDPR datasets and models (aka "gelectra-base-germanquad", "gbert-base-germandpr")
• FARM
• Haystack

Get in touch: Twitter | LinkedIn | Discord | GitHub Discussions | Website

By the way: we're hiring!