language: ko
license: cc-by-4.0
pko-t5-base
pko-t5 λ νκ΅μ΄ μ μ© λ°μ΄ν°λ‘ νμ΅ν t5 v1.1 λͺ¨λΈμ λλ€.
νκ΅μ΄λ₯Ό tokenize νκΈ° μν΄μ sentencepiece λμ OOV κ° μλ BBPE λ₯Ό μ¬μ©νμΌλ©° νκ΅μ΄ λ°μ΄ν° (λ무μν€, μν€νΌλμ, λͺ¨λμλ§λμΉ λ±..) λ₯Ό T5 μ span corruption task λ₯Ό μ¬μ©ν΄μ unsupervised learning λ§ μ μ©νμ¬ νμ΅μ μ§ννμ΅λλ€.
pko-t5 λ₯Ό μ¬μ©νμ€ λλ λμ task μ νμΈνλνμ¬ μ¬μ©νμκΈ° λ°λλλ€.
Usage
transformers μ API λ₯Ό μ¬μ©νμ¬ μ κ·Ό κ°λ₯ν©λλ€. tokenizer λ₯Ό μ¬μ©ν λλ T5Tokenizer
κ° μλλΌ T5TokenizerFast
λ₯Ό μ¬μ©ν΄μ£Όμμμ€. model μ T5ForConditionalGeneration λ₯Ό κ·Έλλ‘ νμ©νμλ©΄ λ©λλ€.
Example
from transformers import T5TokenizerFast, T5ForConditionalGeneration
tokenizer = T5TokenizerFast.from_pretrained('paust/pko-t5-base')
model = T5ForConditionalGeneration.from_pretrained('paust/pko-t5-base')
input_ids = tokenizer(["qa question: λΉμ μ μ΄λ¦μ 무μμΈκ°μ?"]).input_ids
labels = tokenizer(["T5 μ
λλ€."]).input_ids
outputs = model(input_ids=input_ids, labels=labels)
print(f"loss={outputs.loss} logits={outputs.logits}")
Klue νκ° (dev)
Model | ynat (macro F1) | sts (pearsonr/F1) | nli (acc) | ner (entity-level F1) | re (micro F1) | dp (LAS) | mrc (EM/F1) | |
---|---|---|---|---|---|---|---|---|
Baseline | 87.30 | 93.20/86.13 | 89.50 | 86.06 | 71.06 | 87.93 | 75.26/- | |
FT | pko-t5-small (77M) | 86.21 | 77.99/77.01 | 69.20 | 82.60 | 62.95 | 93.15 | 43.81/46.58 |
FT | pko-t5-base (250M) | 87.29 | 90.25/83.43 | 79.73 | 87.80 | 72.94 | 97.28 | 61.53/64.74 |
FT | pko-t5-large (800M) | 87.12 | 92.05/85.24 | 84.96 | 88.18 | 72.26 | 97.60 | 68.01/71.44 |
MT | pko-t5-small | 85.85 | 79.12/77.81 | 66.8 | 81.53 | 67.93 | 91.38 | 44.97/48.07 |
MT | pko-t5-base | 86.86 | 87.61/81.42 | 75.46 | 86.85 | 71.85 | 96.32 | 61.95/65.06 |
MT | pko-t5-large | 87.25 | 91.05/84.58 | 82.16 | 87.63 | 74.78 | 97.33 | 69.18/71.92 |
- FT: μ±κΈνμ€ν¬ νμΈνλ / MT: λ©ν°νμ€ν¬ νμΈνλ
- Baseline: KLUE λ Όλ¬Έμμ μκ°λ dev set μ λν SOTA μ μ
License
PAUSTμμ λ§λ pko-t5λ MIT license νμ 곡κ°λμ΄ μμ΅λλ€.