File size: 3,544 Bytes
00f6699
 
 
554210d
00f6699
 
 
0955864
00f6699
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c35f7a6
00f6699
bb24520
00f6699
 
 
 
 
 
 
67e56de
 
 
 
 
 
 
 
 
 
00f6699
 
 
0955864
 
7f7be89
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
---
language: ko

license: cc-by-4.0
---

# pko-t5-large
[Source Code](https://github.com/paust-team/pko-t5)

pko-t5 λŠ” ν•œκ΅­μ–΄ μ „μš© λ°μ΄ν„°λ‘œ ν•™μŠ΅ν•œ [t5 v1.1 λͺ¨λΈ](https://github.com/google-research/text-to-text-transfer-transformer/blob/84f8bcc14b5f2c03de51bd3587609ba8f6bbd1cd/released_checkpoints.md)μž…λ‹ˆλ‹€.

ν•œκ΅­μ–΄λ₯Ό tokenize ν•˜κΈ° μœ„ν•΄μ„œ sentencepiece λŒ€μ‹  OOV κ°€ μ—†λŠ” BBPE λ₯Ό μ‚¬μš©ν–ˆμœΌλ©° ν•œκ΅­μ–΄ 데이터 (λ‚˜λ¬΄μœ„ν‚€, μœ„ν‚€ν”Όλ””μ•„, λͺ¨λ‘μ˜λ§λ­‰μΉ˜ λ“±..) λ₯Ό T5 의 span corruption task λ₯Ό μ‚¬μš©ν•΄μ„œ unsupervised learning 만 μ μš©ν•˜μ—¬ ν•™μŠ΅μ„ μ§„ν–‰ν–ˆμŠ΅λ‹ˆλ‹€.

pko-t5 λ₯Ό μ‚¬μš©ν•˜μ‹€ λ•ŒλŠ” λŒ€μƒ task 에 νŒŒμΈνŠœλ‹ν•˜μ—¬ μ‚¬μš©ν•˜μ‹œκΈ° λ°”λžλ‹ˆλ‹€.

## Usage
transformers 의 API λ₯Ό μ‚¬μš©ν•˜μ—¬ μ ‘κ·Ό κ°€λŠ₯ν•©λ‹ˆλ‹€. tokenizer λ₯Ό μ‚¬μš©ν• λ•ŒλŠ” `T5Tokenizer` κ°€ μ•„λ‹ˆλΌ `T5TokenizerFast` λ₯Ό μ‚¬μš©ν•΄μ£Όμ‹­μ‹œμ˜€. model 은 T5ForConditionalGeneration λ₯Ό κ·ΈλŒ€λ‘œ ν™œμš©ν•˜μ‹œλ©΄ λ©λ‹ˆλ‹€.

### Example
```python
from transformers import T5TokenizerFast, T5ForConditionalGeneration

tokenizer = T5TokenizerFast.from_pretrained('paust/pko-t5-large')
model = T5ForConditionalGeneration.from_pretrained('paust/pko-t5-large')

input_ids = tokenizer(["qa question: λ‹Ήμ‹ μ˜ 이름은 λ¬΄μ—‡μΈκ°€μš”?"]).input_ids
labels = tokenizer(["T5 μž…λ‹ˆλ‹€."]).input_ids
outputs = model(input_ids=input_ids, labels=labels)

print(f"loss={outputs.loss} logits={outputs.logits}")
```
    

## Klue 평가 (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](https://huggingface.co/paust/pko-t5-small) (77M)  | 86.21           | 77.99/77.01       | 69.20     | 82.60                 | 66.46         | 93.15     | 43.81/46.58 |
| FT  | [pko-t5-base](https://huggingface.co/paust/pko-t5-base) (250M)   | 87.29           | 90.25/83.43       | 79.73     | 87.80                 | 67.23         | 97.28     | 61.53/64.74 |
| FT  | [pko-t5-large](https://huggingface.co/paust/pko-t5-large) (800M) | 87.12           | 92.05/85.24       | 84.96     | **88.18**             | **75.17**     | **97.60** | 68.01/71.44 |
| MT  | pko-t5-small                                                     | 84.54           | 68.50/72/02       | 51.16     | 74.69                 | 66.11         | 80.40     | 43.60/46.28 |
| MT  | pko-t5-base                                                      | 86.89           | 83.96/80.30       | 72.03     | 85.27                 | 66.59         | 95.05     | 61.11/63.94 |
| MT  | pko-t5-large                                                     | 87.57           | 91.93/86.29       | 83.63     | 87.41                 | 71.34         | 96.99     | 70.70/73.72 |

- FT: μ‹±κΈ€νƒœμŠ€ν¬ νŒŒμΈνŠœλ‹ / MT: λ©€ν‹°νƒœμŠ€ν¬ νŒŒμΈνŠœλ‹
- [Baseline](https://arxiv.org/abs/2105.09680): KLUE λ…Όλ¬Έμ—μ„œ μ†Œκ°œλœ dev set 에 λŒ€ν•œ SOTA 점수

## License
[PAUST](https://paust.io)μ—μ„œ λ§Œλ“  pko-t5λŠ” [MIT license](https://github.com/paust-team/pko-t5/blob/main/LICENSE) ν•˜μ— κ³΅κ°œλ˜μ–΄ μžˆμŠ΅λ‹ˆλ‹€.