Skip to content

Latest commit

 

History

History
50 lines (35 loc) · 2.45 KB

short_seq_evaluation_ko.md

File metadata and controls

50 lines (35 loc) · 2.45 KB

Finetune on Short Sequence Dataset

한국어 | English

Details

  • max_seq_length<=512 환경에서의 KoBigBird 성능 평가

  • 5개의 Dataset으로 평가

    • Single Sentence Classification: NSMC
    • Sentence Pair Classification: KLUE-NLI, KLUE-STS
    • Question Answering: Korquad 1.0, KLUE-MRC
  • KLUE-Baseline의 코드를 기반으로 일부 수정하여 학습

    • nsmckorquad 1.0 task 추가
    • transformers==4.11.3에 호환되도록 수정
  • Sequence Classification은 128, Question Answering은 512의 길이로 학습

    • Sparse Attention이 아닌 Full Attention으로 세팅 (아래의 로그가 나오면서 자동으로 Full Attention으로 변경)
    Attention type 'block_sparse' is not possible if sequence_length: 300 <= num global tokens: 2 * config.block_size + min. num sliding tokens: 3 * config.block_size
    + config.num_random_blocks * config.block_size + additional buffer: config.num_random_blocks * config.block_size = 704 with config.block_size = 64, config.num_random_blocks = 3.
    Changing attention type to 'original_full'...
    

Result

NSMC
(acc)
KLUE-NLI
(acc)
KLUE-STS
(pearsonr)
Korquad 1.0
(em/f1)
KLUE MRC
(em/rouge-w)
KoELECTRA-Base-v3 91.13 86.87 93.14 85.66 / 93.94 59.54 / 65.64
KLUE-RoBERTa-Base 91.16 86.30 92.91 85.35 / 94.53 69.56 / 74.64
KoBigBird-BERT-Base 91.18 87.17 92.61 87.08 / 94.71 70.33 / 75.34
  • KLUE, Korquad 1.0 모두 dev set으로 평가
  • KoELECTRA-Base-v3KLUE-RoBERTa-Base의 KLUE dataset 관련 점수는 KLUE PaperA. Dev Set Results에서 참고

Reference