Kaggle Google QA Labeling Competition

Here is solution of 41th place of the Kaggle Google QA Labeling competition

1. Install the package
   pip install -e .

2. Download NER-Model from Kaggle
   https://www.kaggle.com/alexeykarnachev/google-qa-ner

3. Prepare dataset
   python scripts/prepare_dataset.py --seed=228 --train_df_file=data/google-quest-challenge/train.csv --test_df_file=data/google-quest-challenge/test.csv --tokenizer_cls=RobertaTokenizer --tokenizer_path=roberta-large --n_splits=7 --datasets_root=data/datasets/ --crop_strategies=both --dataset_cls=BiDataset --process_math --ner_model_dir=data/ner/code/bert_base_cased
   
   --ner_model_dir is a path to downloaded NER model (from previous step)

4. Run experiment training
   cd scripts
python run_encoder_experiment.py --config_path=../configs/base_config.yaml

5. Wait the training process end ...
   

6. Archive the experiment directory
   cd experiments
tar zcvf <EXPERIMENT_DIR>.tar.gz <EXPERIMENT_DIR>
   

7. Send it to your kaggle datasets storage
   

8. Now, you can inference the model in a kernel
   https://www.kaggle.com/alexeykarnachev/kernel1864bcfc13
   For this, attach the following datasets to the kernel:
   https://www.kaggle.com/alexeykarnachev/kaggle-google-qa-labeling (this package)
   https://www.kaggle.com/alexeykarnachev/google-qa-ner (NER model)
   https://www.kaggle.com/alexeykarnachev/transformersdependencies (transformers lib and dependencies)
   
   Also, attach trained experiment to the kernel
   
   Uncomment all lines in the kernel and replace the EXPERIMENT_NAME placeholder with your experiment name