Dacon-korean-NLI

데이콘 한국어 문장 분류 경진대회에서 PRIVATE Score 0.88415로 23위 (상위 4.9%)를 달성하였습니다.

https://dacon.io/competitions/official/235875/overview/description

• Task : NLI (Natural Language Inference)
• Model : klue/roberta-large 30개 앙상블 (hard voting)
• Data : klue - nli dataset ( https://klue-benchmark.com/tasks/68/data/download )
• GPU : A6000 x 2 (96GB), A5000 x 2 (48GB)

• code
  • data
    • data_aug.tsv
    • train_data.tsv
    • test_data.tsv
  • train.py
  • classify.py
  • utils.py
  • bert_dataset.py

Train

$ python train.py --model_fn [SAVE_PATH] --train_fn [TRAIN_DATA_PATH] --pretrained_model_name [PRETRAINED_MODEL_NAME] 
--valid_ratio [RATIO_OF_VALIDATION_DATA] --batch_size_per_device [BATCH_SIZE] --lr [LEARNING_RATE] --n_epochs [EPOCHS]
--weight_decay [WEIGHT_DECAY] --warmup_ratio [WARM_UP_RATIO] --max_length [MAX_LENGTH] --amp [AUTOMATIC_MIXED_PRECISION]
--drop_out_p [DROP_OUT] --attention_drop_out_p [ATTENTION_DROP_OUT]

Classify

$ cat ./data/test_data.tsv | awk -F '\t' '{print $2, $3}' | python classify.py