emilnuutinen/finsquad

Experiments on fine-tuning a Finnish SQuAD2.0 model on a machine translated Finnish SQuAD2.0 dataset

Fine-tune FinBERT for Finnish SQuAD 2.0

Training scripts are based on Hugginface examples

Results

• Fine-tuned model: https://huggingface.co/TurkuNLP/bert-base-finnish-cased-squad2
• Finnish SQuAD2 dataset: https://huggingface.co/datasets/TurkuNLP/squad_v2_fi

Training examples

Train Finnish BERT-base

python run_qa.py \
  --model_name_or_path bert-base-finnish-cased-v1/ \
  --dataset_name squad_fi.py \
  --version_2_with_negative \
  --overwrite_cache \
  --do_train \
  --do_eval \
  --per_device_train_batch_size 12 \
  --learning_rate 3e-5 \
  --num_train_epochs 2 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --output_dir tmp/bert-base-finnish-cased-squad2/

Only evaluation

python run_qa.py \
  --model_name_or_path bert-base-finnish-cased-v1/ \
  --dataset_name squad_fi.py \
  --version_2_with_negative \
  --overwrite_cache \
  --do_eval \
  --output_dir tmp/bert-base-finnish-cased-squad2_eval/

Train English BERT-base

python run_qa.py \
  --model_name_or_path bert-base-uncased \
  --dataset_name squad_v2 \
  --version_2_with_negative \
  --overwrite_cache \
  --do_train \
  --do_eval \
  --per_device_train_batch_size 12 \
  --learning_rate 3e-5 \
  --num_train_epochs 2 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --output_dir tmp/bert-base-uncased-squad2/