deberta-v3-base-korean

Model Details

DeBERTa는 Disentangled Attention과 Enhanced Masked Language Model을 통해 BERT의 성능을 향상시킨 모델입니다. 그중 DeBERTa V3은 ELECTRA-Style Pre-Training에 Gradient-Disentangled Embedding Sharing을 적용사여 DeBERTA를 개선했습니다.

이 연구는 구글의 TPU Research Cloud(TRC)를 통해 지원받은 Cloud TPU로 학습되었습니다.

How to Get Started with the Model

from transformers import AutoTokenizer, DebertaV2ForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("team-lucid/deberta-v3-base-korean")
model = DebertaV2ForSequenceClassification.from_pretrained("team-lucid/deberta-v3-base-korean")

inputs = tokenizer("안녕, 세상!", return_tensors="pt")
outputs = model(**inputs)

Evaluation

	Backbone Parameters(M)	NSMC (acc)	PAWS (acc)	KorNLI (acc)	KorSTS (spearman)	Question Pair (acc)
DistilKoBERT	22M	88.41	62.55	70.55	73.21	92.48
KoBERT	85M	89.63	80.65	79.00	79.64	93.93
XLM-Roberta-Base	85M	89.49	82.95	79.92	79.09	93.53
KcBERT-Base	85M	89.62	66.95	74.85	75.57	93.93
KcBERT-Large	302M	90.68	70.15	76.99	77.49	94.06
KoELECTRA-Small-v3	9.4M	89.36	77.45	78.60	80.79	94.85
KoELECTRA-Base-v3	85M	90.63	84.45	82.24	85.53	95.25
Ours
DeBERTa-xsmall	22M	91.21	84.40	82.13	83.90	95.38
DeBERTa-small	43M	91.34	83.90	81.61	82.97	94.98
DeBERTa-base	86M	91.22	85.5	82.81	84.46	95.77

* 다른 모델의 결과는 KcBERT-Finetune 과 KoELECTRA를 참고했으며, Hyperparameter 역시 다른 모델과 유사하게 설정습니다.

Model Memory Requirements

dtype	Largest Layer or Residual Group	Total Size	Training using Adam
float32	187.79 MB	513.77 MB	2.01 GB
float16/bfloat16	93.9 MB	256.88 MB	1.0 GB
int8	46.95 MB	128.44 MB	513.77 MB
int4	23.47 MB	64.22 MB	256.88 MB