KoBigBird-KoBart-News-Summarization

This model is a fine-tuned version of noahkim/KoBigBird-KoBart-News-Summarization on the daekeun-ml/naver-news-summarization-ko

Model description

<<20221110 Commit>>

<<KoBigBird-KoBart-News-Summarization ๋ชจ๋ธ ์„ค๋ช…>>

๋‹ค์ค‘๋ฌธ์„œ์š”์•ฝ(Multi-Document-Summarization) Task๋ฅผ ์œ„ํ•ด์„œ KoBigBird ๋ชจ๋ธ์„ Encoder-Decoder๋ชจ๋ธ์„ ๋งŒ๋“ค์–ด์„œ ํ•™์Šต์„ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. KoBigBird๋ฅผ Decoder๋กœ ์“ฐ๋ ค๊ณ  ํ–ˆ์œผ๋‚˜ ์˜ค๋ฅ˜๊ฐ€ ์ƒ๊ฒจ์„œ ์š”์•ฝ์— ํŠนํ™”๋œ KoBART์˜ Decoder๋ฅผ ํ™œ์šฉํ•ด์„œ ๋ชจ๋ธ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.

ํ”„๋กœ์ ํŠธ์šฉ์œผ๋กœ ๋‰ด์Šค ์š”์•ฝ ๋ชจ๋ธ ํŠนํ™”๋œ ๋ชจ๋ธ์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด ๊ธฐ์กด์— ๋งŒ๋“ค์—ˆ๋˜ KoBigBird-KoBart-News-Summarization ๋ชจ๋ธ์— ์ถ”๊ฐ€์ ์œผ๋กœ daekeun-ml๋‹˜์ด ์ œ๊ณตํ•ด์ฃผ์‹  naver-news-summarization-ko ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ํŒŒ์ธํŠœ๋‹ ํ–ˆ์Šต๋‹ˆ๋‹ค.

ํ˜„์žฌ AI-HUB์—์„œ ์ œ๊ณตํ•˜๋Š” ์š”์•ฝ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ ํ•™์Šต ์ง„ํ–‰ ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์ง€์†์ ์œผ๋กœ ๋ฐœ์ „์‹œ์ผœ ์ข‹์€ ์„ฑ๋Šฅ์˜ ๋ชจ๋ธ์„ ๊ตฌํ˜„ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

์‹คํ–‰ํ™˜๊ฒฝ

  • Google Colab Pro
  • CPU : Intel(R) Xeon(R) CPU @ 2.20GHz
  • GPU : A100-SXM4-40GB

# Python Code
from transformers import AutoTokenizer
from transformers import AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("noahkim/KoT5_news_summarization")
model = AutoModelForSeq2SeqLM.from_pretrained("noahkim/KoT5_news_summarization")

The following hyperparameters were used during training:

  • learning_rate: 2e-05
  • train_batch_size: 16
  • eval_batch_size: 16
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • num_epochs: 4
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss
4.0748 1.0 1388 4.3067
3.8457 2.0 2776 4.2039
3.7459 3.0 4164 4.1433
3.6773 4.0 5552 4.1236

Framework versions

  • Transformers 4.24.0
  • Pytorch 1.12.1+cu113
  • Datasets 2.6.1
  • Tokenizers 0.13.2
Downloads last month
22
Inference Examples
Inference API (serverless) has been turned off for this model.