question-answering

DistilBERT with a second step of distillation

Model description

This model replicates the "DistilBERT (D)" model from Table 2 of the DistilBERT paper. In this approach, a DistilBERT student is fine-tuned on SQuAD v1.1, but with a BERT model (also fine-tuned on SQuAD v1.1) acting as a teacher for a second step of task-specific distillation.

In this version, the following pre-trained models were used:

Training data

This model was trained on the SQuAD v1.1 dataset which can be obtained from the datasets library as follows:

from datasets import load_dataset
squad = load_dataset('squad')

Training procedure

Eval results

Exact Match F1
DistilBERT paper 79.1 86.9
Ours 78.4 86.5

The scores were calculated using the squad metric from datasets.

BibTeX entry and citation info

@misc{sanh2020distilbert,
      title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter}, 
      author={Victor Sanh and Lysandre Debut and Julien Chaumond and Thomas Wolf},
      year={2020},
      eprint={1910.01108},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}