BERT Base Uncased Finetuned on TriviaQA

The BERT (Base) model is finetuned on the TriviaQA dataset using a modified version of the run_squad.py legacy script in Transformers. The script is provided in this repository.

$ cd ~/projects/transformers/examples/legacy/question-answering
$ mkdir bert_base_uncased_finetuned_triviaqa
python run_triviaqa.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/triviaqa/squad-triviaqa-wikipedia-train.json" \
        --predict_file "~/projects//data/triviaqa/squad-triviaqa-wikipedia-dev.json" \
        --output_dir "./bert_base_uncased_finetuned_triviaqa" \
        --save_steps 50000

Results:

{'exact': 55.57530864197531, 'f1': 61.37345358329793, 'total': 10125, 'HasAns_exact': 55.57530864197531, 'HasAns_f1': 61.37345358329793, 'HasAns_total': 10125, 'best_exact': 55.57530864197531, 'best_exact_thresh': 0.0, 'best_f1': 61.37345358329793, 'best_f1_thresh': 0.0}