cosmoquester/bart-ko-small

Pretrained BART in Korean

This is pretrained BART model with multiple Korean Datasets.

I used multiple datasets for generalizing the model for both colloquial and written texts.

The training is supported by TPU Research Cloud program.

The script which is used to pre-train model is here.

When you use the reference API, you must wrap the sentence with [BOS] and [EOS] like below example.

[BOS] 안녕하세요? 반가워요~~ [EOS]

You can also test mask filling performance using [MASK] token like this.

[BOS] [MASK] 먹었어? [EOS]

Benchmark

Dataset	KLUE NLI dev	NSMC test	QuestionPair test	KLUE TC dev		KLUE STS dev			KorSTS dev			HateSpeech dev
Metric	Acc	Acc	Acc	Acc	F1	F1	Pearson	Spearman	F1	Pearson	Spearman	Bias Acc	Hate Acc
Score	0.639	0.8721	0.905	0.8551	0.8515	0.7406	0.7593	0.7551	0.7897	0.7269	0.7037	0.8068	0.5966

• The performance was measured using the notebooks here with colab.

Used Datasets

모두의 말뭉치

• 일상 대화 말뭉치 2020
• 구어 말뭉치
• 문어 말뭉치
• 신문 말뭉치

AIhub

• 개방데이터 전문분야말뭉치
• 개방데이터 한국어대화요약
• 개방데이터 감성 대화 말뭉치
• 개방데이터 한국어 음성
• 개방데이터 한국어 SNS

세종 말뭉치