/kor_pratrain_LM

https://ailabs.enliple.com/

Primary LanguagePythonApache License 2.0Apache-2.0

Pre-Trained Korean Language Model

NLP 발전을 위해 한글 Corpus로 Pre-train한 Language Model을 공개합니다.

Large Model의 경우 Fine-Tuning Step에서도 많은 Computational resource가 필요하기 때문에 고사양 Machine이 없을 시 Fine-Tuning이 어렵습니다. 이에 따라 Benchmark를 진행한 3가지 Task(KorQuAD1.0, KorNLI, KorSTS)에 대한 Fine-Tuning Model도 공개합니다.

Pre-train Corpus

  • Small: 한국어 Wikipedia
  • V1: 한국어 Wikipedia + News (88M Sentences)
  • V2: 한국어 Wikipedia + News (174M Sentences)

Model Detail

  • Masking Strategy: Dynamic Masking(RoBERTa) + n-gram Masking(ALBERT)
  • Additional Task: SOP(Sentence Order Prediction)(ALBERT)
  • Optimizer:
  • Scheduler:
  • Mixed-Precision Opt Level "O1" (Nvidia Apex)
  • Hyper-parameters
Hyper-parameter Small Model Large Model
Number of layers 12 24
Hidden Size 256 1024
FFN inner hidden size 1024 4048
Attention heads 8 16
Attention head size 32 64
Mask percent 15 15
Learning Rate 0.0001 0.00125
Warmup Proportion 0.1 0.0125
Attention Dropout 0.1 0.1
Dropout 0.1 0.1
Batch Size 256 2048
Train Steps 500k 125k(V1) 250k(V2)

Model Benchmark

KorQuAD1.0 (EM/F1) KorNLI (acc) KorSTS (spearman)
multilingual-BERT (Base Size) 70.42/90.25 76.33 77.90
KoBERT (Base Size) 52.81/80.27 79.00 79.64
KoELECTRA (Base Size) 61.10/89.59 80.85 83.21
HanBERT (Base Size) 78.74 / 92.02 80.89 83.33
Ours (Small Size) 78.98/88.20 74.67 74.53
Ours (Large Size) 85.61/93.89 81.68 83.90
  • Fine-tuning Setting (Ours Model)
    • Optimizer: Adam
    • Scheduler: LinearWarmup
    • Mixed Precision Opt Level "O2"
    • KorQuAD1.0
      • lr: 5e-5
      • epochs: 4
      • batch size: 16
    • KorNLI
      • lr: 2e-5
      • epochs: 3
      • batch size: 32
    • KorSTS
      • Fine-tuning Step에서 분산은 상당히 클 수 있으므로 상대적으로 Dataset의 크기가 작은 KorSTS Task는 Random Seed에 대해 Grid Search({1~10})를 사용하여 가장 성능이 좋은 Model을 사용하였습니다. (Reference)
      • lr: 3e-5
      • epochs: 10
      • batch size: 16(Large) 32(Small)
      • best random seed: 9(Large) 7(Small)

Example Scripts

KorQuAD1.0

  • Train
python3 run_qa.py \
  --checkpoint $MODEL_FILE \
  --config_file $CONFIG_FILE \
  --vocab_file $VOCAB_FILE \
  --train_file data/korquad/KorQuAD_v1.0_train.json \
  --max_seq_length 512 \
  --doc_stride 128 \
  --max_query_length 64 \
  --max_answer_length 30 \
  --per_gpu_train_batch_size 16 \
  --learning_rate 5e-5 \
  --num_train_epochs 4.0 \
  --adam_epsilon 1e-6 \
  --warmup_proportion 0.1
  • Eval
python3 eval_qa.py \
  --checkpoint $MODEL_FILE \
  --config_file $CONFIG_FILE \
  --vocab_file $VOCAB_FILE \
  --predict_file data/korquad/KorQuAD_v1.0_dev.json \
  --max_seq_length 512 \
  --doc_stride 64 \
  --max_query_length 64 \
  --max_answer_length 30 \
  --batch_size 16 \
  --n_best_size 20

KorNLI

  • Train
python3 run_classifier.py \
  --data_dir data/kornli \
  --task_name kornli \
  --config_file $CONFIG_FILE \
  --vocab_file $VOCAB_FILE \
  --checkpoint $MODEL_FILE \
  --do_eval \
  --max_seq_length 128 \
  --train_batch_size 32 \
  --eval_batch_size 32 \
  --learning_rate 3e-5 \
  --num_train_epochs 3.0 \
  --warmup_proportion 0.1
  • Eval
python3 eval_classifier.py \
  --data_dir data/kornli \
  --task_name kornli \
  --config_file $CONFIG_FILE \
  --vocab_file $VOCAB_FILE \
  --checkpoint $MODEL_FILE \
  --max_seq_length 128 \
  --eval_batch_size 32

KorSTS

  • Train
python3 run_classifier.py \
  --data_dir data/korsts \
  --task_name korsts \
  --config_file $CONFIG_FILE \
  --vocab_file $VOCAB_FILE \
  --checkpoint $MODEL_FILE \
  --do_eval
  --max_seq_length 128 \
  --train_batch_size 16 \
  --eval_batch_size 32 \
  --learning_rate 3e-5 \
  --num_train_epochs 10.0 \
  --warmup_proportion 0.1
  • Eval
python3 eval_classifier.py \
  --data_dir data/korsts \
  --task_name korsts \
  --config_file $CONFIG_FILE \
  --vocab_file $VOCAB_FILE \
  --checkpoint $MODEL_FILE \
  --max_seq_length 128 \
  --eval_batch_size 32

Acknowledgement

본 연구는 과학기술정보통신부 및 정보통신산업진흥원의 ‘고성능 컴퓨팅 지원’ 사업으로부터 지원받아 수행하였음
Following(or This research) was results of a study on the "HPC Support" Project, supported by the ‘Ministry of Science and ICT’ and NIPA.

Reference

  • 추가적으로 궁금하신점은 해당 repo의 issue를 등록해주시거나 ekjeon@enliple.com으로 메일 주시면 답변 드리겠습니다.