README.md

metadata

language:
  - ko

NewsKoT5

The training data for this T5 model consists of Korean news articles (29GB). However, the performance has not been fine-tuned through the use of small batches and a limited number of training steps, so it may not be fully optimized.

Quick tour

from transformers import AutoTokenizer, T5ForConditionalGeneration
  
tokenizer = AutoTokenizer.from_pretrained("BM-K/NewsKoT5-small")
model = T5ForConditionalGeneration.from_pretrained("BM-K/NewsKoT5-small")

input_ids = tokenizer("한국형발사체 누리호가 실용급 <extra_id_0> 발사체로서 ‘데뷔’를 성공적으로 <extra_id_1>", return_tensors="pt").input_ids
labels = tokenizer("<extra_id_0> 위성 <extra_id_1> 마쳤다 <extra_id_2>", return_tensors="pt").input_ids

outputs = model(input_ids=input_ids,
                labels=labels)

News Summarization Performance (F1-score)

After restoring the model's tokenized output to the original text, Rouge performance was evaluated by comparing it to the reference and hypothesis tokenized using mecab.

• Dacon 한국어 문서 생성요약 AI 경진대회 Dataset
  • Training: 29,432
  • Validation: 7,358
  • Test: 9,182

	#Param	rouge-1	rouge-2	rouge-l
pko-t5-small	95M	51.48	33.18	44.96
NewsT5-small	61M	52.15	33.59	45.41

• AI-Hub 문서요약 텍스트 Dataset
  • Training: 245,626
  • Validation: 20,296
  • Test: 9,931

	#Param	rouge-1	rouge-2	rouge-l
pko-t5-small	95M	53.44	34.03	45.36
NewsT5-small	61M	53.74	34.27	45.52

• pko-t5-small