Beomi/KcBERT-Finetune

KcBERT/KcELECTRA Fine Tune Benchmarks code (forked from https://github.com/monologg/KoELECTRA/tree/master/finetune)

한국어 | English

Finetuning (Benchmark on subtask)

• KoELECTRA Finetune Code를 Fork하여 제작
• Single GPU 사용

Requirements

• python 3.9
• pip reqs

torch<=1.11
transformers<5
seqeval
fastprogress
attrdict

How to Run

$ python3 run_seq_cls.py --task {$TASK_NAME} --config_file {$CONFIG_FILE}

KcBERT

# Base: kcbert-base.json / Large: kcbert-large.json
$ python3 run_seq_cls.py --task nsmc --config_file kcbert-base.json
$ python3 run_seq_cls.py --task kornli --config_file kcbert-base.json
$ python3 run_seq_cls.py --task paws --config_file kcbert-base.json
$ python3 run_seq_cls.py --task question-pair --config_file kcbert-base.json
$ python3 run_seq_cls.py --task korsts --config_file kcbert-base.json
$ python3 run_ner.py --task naver-ner --config_file kcbert-base.json
$ python3 run_squad.py --task korquad --config_file kcbert-base.json

KcBERT-v2022

# run_all => bash run_all_kcbert_v2022.sh

python3 run_seq_cls.py --task nsmc --config_file kcbert-base-v2022.json
python3 run_seq_cls.py --task kornli --config_file kcbert-base-v2022.json
python3 run_seq_cls.py --task paws --config_file kcbert-base-v2022.json
python3 run_seq_cls.py --task question-pair --config_file kcbert-base-v2022.json
python3 run_seq_cls.py --task korsts --config_file kcbert-base-v2022.json
python3 run_ner.py --task naver-ner --config_file kcbert-base-v2022.json
python3 run_squad.py --task korquad --config_file kcbert-base-v2022.json

KcELECTRA

Base Command

# Base: kcelectra-base.json
$ python3 run_seq_cls.py --task nsmc --config_file kcelectra-base-v1.json
$ python3 run_seq_cls.py --task kornli --config_file kcelectra-base-v1.json
$ python3 run_seq_cls.py --task paws --config_file kcelectra-base-v1.json
$ python3 run_seq_cls.py --task question-pair --config_file kcelectra-base-v1.json
$ python3 run_seq_cls.py --task korsts --config_file kcelectra-base-v1.json
$ python3 run_ner.py --task naver-ner --config_file kcelectra-base-v1.json
$ python3 run_squad.py --task korquad --config_file kcelectra-base-v1.json

KcELECTRA-v2022-dev

Base Command

export CUDA_VISIBLE_DEVICES=1
python3 run_seq_cls.py --task nsmc --config_file kcelectra-v2022-dev.json
python3 run_seq_cls.py --task kornli --config_file kcelectra-v2022-dev.json
python3 run_seq_cls.py --task paws --config_file kcelectra-v2022-dev.json
python3 run_seq_cls.py --task question-pair --config_file kcelectra-v2022-dev.json
python3 run_seq_cls.py --task korsts --config_file kcelectra-v2022-dev.json
python3 run_ner.py --task naver-ner --config_file kcelectra-v2022-dev.json
python3 run_squad.py --task korquad --config_file kcelectra-v2022-dev.json

KcELECTRA-v2022-v2-namu

Base Command

CUDA_VISIBLE_DEVICES=0 python3 run_seq_cls.py --task nsmc --config_file kcelectra-v2022-v2-300k-discriminator.json
CUDA_VISIBLE_DEVICES=0 python3 run_seq_cls.py --task kornli --config_file kcelectra-v2022-v2-300k-discriminator.json

CUDA_VISIBLE_DEVICES=1 python3 run_seq_cls.py --task paws --config_file kcelectra-v2022-v2-300k-discriminator.json
CUDA_VISIBLE_DEVICES=1 python3 run_seq_cls.py --task question-pair --config_file kcelectra-v2022-v2-300k-discriminator.json

CUDA_VISIBLE_DEVICES=2 python3 run_seq_cls.py --task korsts --config_file kcelectra-v2022-v2-300k-discriminator.json
CUDA_VISIBLE_DEVICES=2 python3 run_ner.py --task naver-ner --config_file kcelectra-v2022-v2-300k-discriminator.json

CUDA_VISIBLE_DEVICES=3 python3 run_squad.py --task korquad --config_file kcelectra-v2022-v2-300k-discriminator.json

exKcBERT-paws

Command

# Base: exkcbert-paws.json
# $ python3 run_seq_cls.py --task nsmc --config_file exkcbert-paws.json
# $ python3 run_seq_cls.py --task kornli --config_file exkcbert-paws.json
$ python3 run_seq_cls.py --task paws --config_file exkcbert-paws.json
# $ python3 run_seq_cls.py --task question-pair --config_file exkcbert-paws.json
# $ python3 run_seq_cls.py --task korsts --config_file exkcbert-paws.json
# $ python3 run_ner.py --task naver-ner --config_file exkcbert-paws.json
$ python3 run_squad.py --task korquad --config_file exkcbert-paws.json

exKcBERT-paws-extonly

Command

# Base: exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task nsmc --config_file exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task kornli --config_file exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task paws --config_file exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task question-pair --config_file exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task korsts --config_file exkcbert-paws-extonly.json
$ python3 run_ner.py --task naver-ner --config_file exkcbert-paws-extonly.json
$ python3 run_squad.py --task korquad --config_file exkcbert-paws-extonly.json

exKcBERT-paws-extonly

Command

# Base: exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task nsmc --config_file exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task kornli --config_file exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task paws --config_file exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task question-pair --config_file exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task korsts --config_file exkcbert-paws-extonly.json
$ python3 run_ner.py --task naver-ner --config_file exkcbert-paws-extonly.json
$ python3 run_squad.py --task korquad --config_file exkcbert-paws-extonly.json

exKcBERT-kowiki-extonly

Command

# Base: exkcbert-paws-extonly.json
$ python3 run_seq_cls.py --task nsmc --config_file exkcbert-kowiki.json
$ python3 run_seq_cls.py --task kornli --config_file exkcbert-kowiki.json
$ python3 run_seq_cls.py --task paws --config_file exkcbert-kowiki.json
$ python3 run_seq_cls.py --task question-pair --config_file exkcbert-kowiki.json
$ python3 run_seq_cls.py --task korsts --config_file exkcbert-kowiki.json
$ python3 run_ner.py --task naver-ner --config_file exkcbert-kowiki.json
$ python3 run_squad.py --task korquad --config_file exkcbert-kowiki.json

# once
python3 run_seq_cls.py --task nsmc --config_file exkcbert-kowiki.json && \
python3 run_seq_cls.py --task kornli --config_file exkcbert-kowiki.json && \
python3 run_seq_cls.py --task paws --config_file exkcbert-kowiki.json && \
python3 run_seq_cls.py --task question-pair --config_file exkcbert-kowiki.json && \
python3 run_seq_cls.py --task korsts --config_file exkcbert-kowiki.json && \
python3 run_ner.py --task naver-ner --config_file exkcbert-kowiki.json && \
python3 run_squad.py --task korquad --config_file exkcbert-kowiki.json 

Result

Base Model

	Size	NSMC (acc)	Naver NER (F1)	PAWS (acc)	KorNLI (acc)	KorSTS (spearman)	Question Pair (acc)	KorQuaD (Dev) (EM/F1)
exKcBERT-paws	641M	n/a	n/a	TODO	n/a	n/a	n/a	TODO
KcELECTRA-base	475M	91.71	86.90	74.80	81.65	82.65	95.78	70.60 / 90.11
KcBERT-Base	417M	89.62	84.34	66.95	74.85	75.57	93.93	60.25 / 84.39
KcBERT-Large	1.2G	90.68	85.53	70.15	76.99	77.49	94.06	62.16 / 86.64
KoBERT	351M	89.63	86.11	80.65	79.00	79.64	93.93	52.81 / 80.27
XLM-Roberta-Base	1.03G	89.49	86.26	82.95	79.92	79.09	93.53	64.70 / 88.94
HanBERT	614M	90.16	87.31	82.40	80.89	83.33	94.19	78.74 / 92.02
KoELECTRA-Base	423M	90.21	86.87	81.90	80.85	83.21	94.20	61.10 / 89.59
KoELECTRA-Base-v2	423M	89.70	87.02	83.90	80.61	84.30	94.72	84.34 / 92.58
DistilKoBERT	108M	88.41	84.13	62.55	70.55	73.21	92.48	54.12 / 77.80

*HanBERT의 Size는 Bert Model과 Tokenizer DB를 합친 것입니다.

*config의 세팅을 그대로 하여 돌린 결과이며, hyperparameter tuning을 추가적으로 할 시 더 좋은 성능이 나올 수 있습니다.

Reference

• Transformers Examples
• NSMC
• Naver NER Dataset
• PAWS
• KorNLI/KorSTS
• Question Pair
• KorQuad
• KoBERT
• HanBERT
• HanBert Transformers
• KoELECTRA