File size: 2,833 Bytes
51f763f
 
 
 
 
 
 
 
0c146e7
2ca7636
51f763f
 
 
 
 
2ca7636
51f763f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2ca7636
51f763f
 
 
2ca7636
51f763f
 
 
 
 
 
 
 
 
 
 
 
2ca7636
 
 
 
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
52
53
54
55
56
57
58
59
60
61
---
language: ko
license: mit
library_name: transformers
pipeline_tag: text2text-generation
---

# FLAN T5
[Source Code](https://github.com/paust-team/pko-t5/tree/main/pkot5/flan)

FLAN T5λŠ” [paust/pko-t5-large](https://huggingface.co/paust/pko-t5-large) λͺ¨λΈμ„ 기반으둜 λ‹€μ–‘ν•œ νƒœμŠ€ν¬λ₯Ό instruction finetuning을 ν†΅ν•΄μ„œ λ§Œλ“  λͺ¨λΈμž…λ‹ˆλ‹€.

ν˜„μž¬ 계속 Instruction Finetuning 을 μ§„ν–‰ν•˜λ©΄μ„œ 쀑간결과λ₯Ό λͺ¨λΈλ‘œ μ—…λ°μ΄νŠΈν•˜κ³  μžˆμŠ΅λ‹ˆλ‹€.


## ν•™μŠ΅λœ νƒœμŠ€ν¬
| Task name                  | Task type      | 
|----------------------------|----------------|
| NSMC                       | Classification |
| Klue Ynat                  | Classification |
| KorNLI                     | Classification |
| KorSTS                     | Classification |
| QuestionPair               | Classification |
| Klue STS                   | Classification |
| AIHub news Summary         | Summarization  |
| AIHub document Summary     | Summarization  |
| AIHub book Summary         | Summarization  |
| AIHub conversation Summary | Summarization  |
| AIHub ko-to-en             | Translation    |
| AIHub ko-to-en Expert      | Translation    |
| AIHub ko-to-en Tech        | Translation    |
| AIHub ko-to-en social      | Translation    |
| AIHub ko-to-jp             | Translation    |
| AIHub ko-to-cn Tech        | Translation    |
| AIHub Translation Corpus   | Translation    |
| korquad                    | QA             |
| Klue MRC                   | QA             |
| AIHub mindslab's MRC       | QA             |


## λͺ¨λΈ
- [Hugginface 링크](https://huggingface.co/paust/pko-flan-t5-large)


## μ‚¬μš© μ˜ˆμ‹œ
```python
from transformers import T5ForConditionalGeneration, T5TokenizerFast

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

prompt = """μ„œμšΈνŠΉλ³„μ‹œ(μ„œμšΈη‰Ήεˆ₯εΈ‚, μ˜μ–΄: Seoul Metropolitan Government)λŠ” λŒ€ν•œλ―Όκ΅­ μˆ˜λ„μ΄μž μ΅œλŒ€ λ„μ‹œμ΄λ‹€. μ„ μ‚¬μ‹œλŒ€λΆ€ν„° μ‚¬λžŒμ΄ κ±°μ£Όν•˜μ˜€μœΌλ‚˜ λ³Έ μ—­μ‚¬λŠ” 백제 첫 μˆ˜λ„ μœ„λ‘€μ„±μ„ μ‹œμ΄ˆλ‘œ ν•œλ‹€. μ‚Όκ΅­μ‹œλŒ€μ—λŠ” μ „λž΅μ  μš”μΆ©μ§€λ‘œμ„œ 고ꡬ렀, 백제, 신라가 λ²ˆκ°ˆμ•„ μ°¨μ§€ν•˜μ˜€μœΌλ©°, κ³ λ € μ‹œλŒ€μ—λŠ” μ™•μ‹€μ˜ 별ꢁ이 μ„Έμ›Œμ§„ 남경(南京)으둜 μ΄λ¦„ν•˜μ˜€λ‹€.
ν•œκ΅­μ˜ μˆ˜λ„λŠ” μ–΄λ””μž…λ‹ˆκΉŒ?"""
input_ids = tokenizer(prompt, add_special_tokens=True, return_tensors='pt').input_ids
output_ids = model.generate(input_ids=input_ids.cuda(), max_new_tokens=32, num_beams=12)
text = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
print(text)  # μ„œμšΈνŠΉλ³„μ‹œ
```

## 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‹€.