JeBERT_ko_je_v2 / README.md
kompactss's picture
Update README.md
db6d1e3
|
raw
history blame
567 Bytes
metadata
license: afl-3.0

๐ŸŠ ์ œ์ฃผ ๋ฐฉ์–ธ ๋ฒˆ์—ญ ๋ชจ๋ธ ๐ŸŠ

  • ํ‘œ์ค€์–ด -> ์ œ์ฃผ์–ด
  • Made by. ๊ตฌ๋ฆ„ ์ž์—ฐ์–ด์ฒ˜๋ฆฌ ๊ณผ์ • 3๊ธฐ 3์กฐ!!

1. Seq2Seq Transformer Model

  • encoder : Bert
  • decoder : Bert
  • Tokenizer : WordPiece Tokenizer

2. Dataset

  • Jit Dataset
  • AI HUB(+์•„๋ž˜์•„ ๋ฌธ์ž)_v2

3. Hyper Parameters

  • Epoch : 10 epochs(best at 7 epoch)
  • Random Seed : 42
  • Learning Rate : 5e-5
  • Warm up Ratio : 0.1
  • Batch Size : 32

4. BLEU Score

  • Jit + AI HUB(+์•„๋ž˜์•„ ๋ฌธ์ž) Dataset : 67.6