MobileBERT fine-tuned on SQuAD v2

MobileBERT is a thin version of BERT_LARGE, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks.

This model was fine-tuned from the HuggingFace checkpoint google/mobilebert-uncased on SQuAD2.0.

Details

Fine-tuning

• Python: 3.7.5

• Machine specs:

  CPU: Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz

  Memory: 32 GiB

  GPUs: 2 GeForce GTX 1070, each with 8GiB memory

  GPU driver: 418.87.01, CUDA: 10.1

• script:

  # after install https://github.com/huggingface/transformers
  
  cd examples/question-answering
  mkdir -p data
  
  wget -O data/train-v2.0.json https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json
  
  wget -O data/dev-v2.0.json  https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json
  
  export SQUAD_DIR=`pwd`/data
  
  python run_squad.py \
    --model_type mobilebert \
    --model_name_or_path google/mobilebert-uncased \
    --do_train \
    --do_eval \
    --do_lower_case \
    --version_2_with_negative \
    --train_file $SQUAD_DIR/train-v2.0.json \
    --predict_file $SQUAD_DIR/dev-v2.0.json \
    --per_gpu_train_batch_size 16 \
    --per_gpu_eval_batch_size 16 \
    --learning_rate 4e-5 \
    --num_train_epochs 5.0 \
    --max_seq_length 320 \
    --doc_stride 128 \
    --warmup_steps 1400 \
    --save_steps 2000 \
    --output_dir $SQUAD_DIR/mobilebert-uncased-warmup-squad_v2 2>&1 | tee train-mobilebert-warmup-squad_v2.log
  

It took about 3.5 hours to finish.

Results

Model size: 95M

Note that the above results didn't involve any hyperparameter search.

Example Usage

from transformers import pipeline

qa_pipeline = pipeline(
    "question-answering",
    model="csarron/mobilebert-uncased-squad-v2",
    tokenizer="csarron/mobilebert-uncased-squad-v2"
)

predictions = qa_pipeline({
    'context': "The game was played on February 7, 2016 at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California.",
    'question': "What day was the game played on?"
})

print(predictions)
# output:
# {'score': 0.71434086561203, 'start': 23, 'end': 39, 'answer': 'February 7, 2016'}

Created by Qingqing Cao | GitHub | Twitter

Made with ❤️ in New York.