metadata

language:
  - en
tags:
  - question-answering
license: apache-2.0
datasets:
  - newsqa
metrics:
  - f1
  - exact_match

BERT Base Uncased Finetuned on NewsQA

The BERT (Base) model is finetuned on the NewsQA dataset using a modified version of the run_squad.py legacy script in Transformers. The script is provided in this repository. Examples with noAnswer and badQuestion are not included in the training process.

$ cd ~/projects/transformers/examples/legacy/question-answering
$ mkdir bert_base_uncased_finetuned_newsqa
$ python run_newsqa.py \
    --model_type bert \
    --model_name_or_path "bert-base-uncased" \
    --do_train \
    --do_eval \
    --do_lower_case \
    --num_train_epochs 2 \
    --per_gpu_train_batch_size 8 \
    --per_gpu_eval_batch_size 32 \
    --max_seq_length 384 \
    --max_grad_norm inf \
    --doc_stride 128 \
    --train_file "~/projects/data/newsqa/combined-newsqa-data-v1.json" \
    --predict_file "~/projects/data/newsqa/combined-newsqa-data-v1.json" \
    --output_dir "./bert_base_uncased_finetuned_newsqa" \
    --save_steps 20000

Results:

{'exact': 60.19350380096752, 'f1': 73.29371985128037, 'total': 4341, 'HasAns_exact': 60.19350380096752, 'HasAns_f1': 73.29371985128037, 'HasAns_total': 4341, 'best_exact': 60.19350380096752, 'best_exact_thresh': 0.0, 'best_f1': 73.29371985128037, 'best_f1_thresh': 0.0}

To prepare the database, follow the instructions on the NewsQA repository.

Evaluate the finetuned model:

python run_newsqa.py \
  --model_type bert \
  --model_name_or_path "./bert_large_uncased_finetuned_newsqa/checkpoint-700000" \
  --do_eval \
  --do_lower_case \
  --per_gpu_eval_batch_size 32 \
  --max_seq_length 384 \
  --max_grad_norm inf \
  --doc_stride 128 \
  --predict_file "~/projects/data/newsqa/combined-newsqa-data-v1.json" \
  --output_dir "./bert_large_uncased_finetuned_newsqa_eval"