File size: 7,555 Bytes
077cd3b
 
 
 
 
 
3b57260
7dc1eb9
3b57260
 
 
e647d18
46263fb
dd070bf
3b57260
 
 
 
 
 
 
 
 
7dc1eb9
 
 
3b57260
7dc1eb9
3b57260
 
 
7dc1eb9
 
3b57260
 
 
023ed48
 
021ba4a
3b57260
ed04a7c
 
 
f547b54
92d77fe
a936ddd
09cb4b6
c0bc444
c81ad98
8e628eb
 
 
a936ddd
8e628eb
 
 
09cb4b6
8e628eb
 
 
6974279
c81ad98
3b57260
021ba4a
023ed48
 
085d284
 
8aef4dd
0704ce7
023ed48
 
 
 
 
 
 
 
3b57260
 
 
7dc1eb9
bba68af
7dc1eb9
bba68af
7dc1eb9
bba68af
7dc1eb9
3b57260
 
 
 
7dc1eb9
bba68af
7dc1eb9
bba68af
7dc1eb9
3b57260
 
 
 
 
 
 
e647d18
3b57260
 
 
dd070bf
3b57260
dd070bf
3b57260
7dc1eb9
 
 
7e8de0f
7dc1eb9
3b57260
 
 
7dc1eb9
 
 
 
 
 
 
3b57260
 
 
 
 
 
 
 
 
 
 
 
e647d18
 
3b57260
 
 
 
b78d70a
af980aa
b78d70a
3b57260
 
 
b78d70a
3b57260
 
 
7dc1eb9
3b57260
b77f831
3b57260
a827a28
b77f831
6d821ea
b77f831
6d821ea
3b57260
 
 
7dc1eb9
3b57260
 
 
7dc1eb9
3b57260
 
 
c918f5f
3b57260
7dc1eb9
3b57260
 
7dc1eb9
3b57260
 
7dc1eb9
3b57260
ba09b42
3b57260
 
 
8e628eb
3b57260
 
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
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
---
license: mit
language:
- ko
metrics:
- accuracy
---
# Model Card for KorSciDeBERTa

<!-- Provide a quick summary of what the model is/does. -->

KorSciDeBERTa๋Š” Microsoft DeBERTa ๋ชจ๋ธ์˜ ์•„ํ‚คํ…์ณ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ, ๋…ผ๋ฌธ, ์—ฐ๊ตฌ ๋ณด๊ณ ์„œ, ํŠนํ—ˆ, ๋‰ด์Šค, ํ•œ๊ตญ์–ด ์œ„ํ‚ค ๋ง๋ญ‰์น˜ ์ด 146GB๋ฅผ ์‚ฌ์ „ํ•™์Šตํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

๋งˆ์Šคํ‚น๋œ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋˜๋Š” ๋‹ค์Œ ๋ฌธ์žฅ ์˜ˆ์ธก์— ์‚ฌ์ „ํ•™์Šต ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ณ , ์ถ”๊ฐ€๋กœ ๋ฌธ์žฅ ๋ถ„๋ฅ˜, ๋‹จ์–ด ํ† ํฐ ๋ถ„๋ฅ˜ ๋˜๋Š” ์งˆ์˜์‘๋‹ต๊ณผ ๊ฐ™์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->



- **Developed by:** KISTI
- **Model type:** deberta-v2
- **Language(s) (NLP):** ํ•œ๊ธ€(ko)

### Model Sources

<!-- Provide the basic links for the model. -->

- **Repository 1:** https://huggingface.co/kisti/korscideberta
- **Repository 2:** https://aida.kisti.re.kr/

## Uses

### Downstream Use

### Load Huggingface model directly
<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
1. ํ˜•ํƒœ์†Œ ๋ถ„์„๊ธฐ(Mecab) ๋“ฑ ์„ค์น˜ ํ•„์ˆ˜ - KorSciDeBERTa ํ™˜๊ฒฝ์„ค์น˜+ํŒŒ์ธํŠœ๋‹.pdf

- Mecab ์„ค์น˜ ์ฐธ๊ณ : ๋‹ค์Œ ๋งํฌ์—์„œ '์‚ฌ์šฉ๋ฐฉ๋ฒ•'. https://aida.kisti.re.kr/model/9bbabd2d-6ce8-44cc-b2a3-69578d23970a

2. git clone https://huggingface.co/kisti/korscideberta; cd korscideberta

- **korscideberta-abstractcls.ipynb**

<pre><code>
from tokenization_korscideberta import DebertaV2Tokenizer
from transformers import AutoModelForSequenceClassification


tokenizer = DebertaV2Tokenizer.from_pretrained("kisti/korscideberta")
model = AutoModelForSequenceClassification.from_pretrained("kisti/korscideberta", num_labels=6, hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1)
#model = AutoModelForMaskedLM.from_pretrained("kisti/korscideberta")
''''''
train_metrics = trainer.train().metrics
trainer.save_metrics("train", train_metrics)
trainer.push_to_hub()
  
</code></pre>

### KorSciDeBERTa native code
KorSciDeBERTa ํ™˜๊ฒฝ์„ค์น˜+ํŒŒ์ธํŠœ๋‹.pdf ์ฐธ์กฐ

- ํŒŒ์ธํŠœ๋‹(์›๋ณธ DeBERTa ์ฐธ๊ณ ): https://github.com/microsoft/DeBERTa/tree/master#run-deberta-experiments-from-command-line

- **korscideberta.zip**

<pre><code>
git clone https://huggingface.co/kisti/korscideberta; cd korscideberta; unzip korscideberta.zip -d korscideberta
''''''
cd korscideberta/experiments/glue;ย chmod 777 *.sh;
./mnli.sh

</code></pre>

### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
์ด ๋ชจ๋ธ์€ ์˜๋„์ ์œผ๋กœ ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ์ ๋Œ€์ ์ด๋‚˜ ์†Œ์™ธ๋œ ํ™˜๊ฒฝ์„ ์กฐ์„ฑํ•˜๋Š”๋ฐ ์‚ฌ์šฉ๋˜์–ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค.

์ด ๋ชจ๋ธ์€ '๊ณ ์œ„ํ—˜ ์„ค์ •'์—์„œ ์‚ฌ์šฉ๋  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์‚ฌ๋žŒ์ด๋‚˜ ์‚ฌ๋ฌผ์— ๋Œ€ํ•œ ์ค‘์š”ํ•œ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๊ฒŒ ์„ค๊ณ„๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ ์ถœ๋ ฅ๋ฌผ์€ ์‚ฌ์‹ค์ด ์•„๋‹ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

'๊ณ ์œ„ํ—˜ ์„ค์ •'์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‚ฌํ•ญ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค:

์˜๋ฃŒ/์ •์น˜/๋ฒ•๋ฅ /๊ธˆ์œต ๋ถ„์•ผ์—์„œ์˜ ์‚ฌ์šฉ, ๊ณ ์šฉ/๊ต์œก/์‹ ์šฉ ๋ถ„์•ผ์—์„œ์˜ ์ธ๋ฌผ ํ‰๊ฐ€, ์ž๋™์œผ๋กœ ์ค‘์š”ํ•œ ๊ฒƒ์„ ๊ฒฐ์ •ํ•˜๊ธฐ, (๊ฐ€์งœ)์‚ฌ์‹ค์„ ์ƒ์„ฑํ•˜๊ธฐ, ์‹ ๋ขฐ๋„ ๋†’์€ ์š”์•ฝ๋ฌธ ์ƒ์„ฑ, ํ•ญ์ƒ ์˜ณ์•„์•ผ๋งŒ ํ•˜๋Š” ์˜ˆ์ธก ์ƒ์„ฑ ๋“ฑ.

## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->
์—ฐ๊ตฌ๋ชฉ์ ์œผ๋กœ ์ €์ž‘๊ถŒ ๋ฌธ์ œ๊ฐ€ ์—†๋Š” ๋ง๋ญ‰์น˜ ๋ฐ์ดํ„ฐ๋งŒ์„ ์‚ฌ์šฉํ•˜์˜€์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์˜ ์‚ฌ์šฉ์ž๋Š” ์•„๋ž˜์˜ ์œ„ํ—˜ ์š”์ธ๋“ค์„ ์ธ์‹ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

์‚ฌ์šฉ๋œ ๋ง๋ญ‰์น˜๋Š” ๋Œ€๋ถ€๋ถ„ ์ค‘๋ฆฝ์ ์ธ ์„ฑ๊ฒฉ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š”๋ฐ๋„ ๋ถˆ๊ตฌํ•˜๊ณ , ์–ธ์–ด ๋ชจ๋ธ์˜ ํŠน์„ฑ์ƒ ์•„๋ž˜์™€ ๊ฐ™์€ ์œค๋ฆฌ ๊ด€๋ จ ์š”์†Œ๋ฅผ ์ผ๋ถ€ ํฌํ•จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:

ํŠน์ • ๊ด€์ ์— ๋Œ€ํ•œ ๊ณผ๋Œ€/๊ณผ์†Œ ํ‘œํ˜„, ๊ณ ์ • ๊ด€๋…, ๊ฐœ์ธ ์ •๋ณด, ์ฆ์˜ค/๋ชจ์š• ๋˜๋Š” ํญ๋ ฅ์ ์ธ ์–ธ์–ด, ์ฐจ๋ณ„์ ์ด๊ฑฐ๋‚˜ ํŽธ๊ฒฌ์ ์ธ ์–ธ์–ด, ๊ด€๋ จ์ด ์—†๊ฑฐ๋‚˜ ๋ฐ˜๋ณต์ ์ธ ์ถœ๋ ฅ ์ƒ์„ฑ ๋“ฑ.


## Training Details

### Training Data

<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
๋…ผ๋ฌธ, ์—ฐ๊ตฌ ๋ณด๊ณ ์„œ, ํŠนํ—ˆ, ๋‰ด์Šค, ํ•œ๊ตญ์–ด ์œ„ํ‚ค ๋ง๋ญ‰์น˜ ์ด 146GB

### Training Procedure 

KISTI HPC NVIDIA A100 80G GPU 24EA์—์„œ 2.5๊ฐœ์›”๋™์•ˆ 1,600,000 ์Šคํ… ํ•™์Šต

#### Preprocessing

- ๊ณผํ•™๊ธฐ์ˆ ๋ถ„์•ผ ํ† ํฌ๋‚˜์ด์ € (KorSci Tokenizer)
- ๋ณธ ์‚ฌ์ „ํ•™์Šต ๋ชจ๋ธ์—์„œ ์‚ฌ์šฉ๋œ ์ฝ”ํผ์Šค๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ช…์‚ฌ ๋ฐ ๋ณตํ•ฉ๋ช…์‚ฌ ์•ฝ 600๋งŒ๊ฐœ์˜ ์‚ฌ์šฉ์ž์‚ฌ์ „์ด ์ถ”๊ฐ€๋œ [Mecab-ko Tokenizer](https://bitbucket.org/eunjeon/mecab-ko/src/master/)์™€ ๊ธฐ์กด SentencePiece-BPE๊ฐ€ ๋ณ‘ํ•ฉ๋˜์–ด์ง„ ํ† ํฌ๋‚˜์ด์ €๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ง๋ญ‰์น˜๋ฅผ ์ „์ฒ˜๋ฆฌํ•˜์˜€์Šต๋‹ˆ๋‹ค.
 - Total 128,100 words
 - Included special tokens ( < unk >, < cls >, < s >, < mask > )
 - File name : spm.model, vocab.txt

#### Training Hyperparameters

- **model_size:** base
- **num_train_steps:** 1,600,000
- **train_batch_size:** 4,096 * 4 accumulative update = 16,384
- **learning_rate:** 1e-4
- **max_seq_length:** 512
- **vocab_size:** 128,100
- **Training regime:** fp16 mixed precision <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->


## Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

### Testing Data, Factors & Metrics

#### Testing Data

<!-- This should link to a Data Card if possible. -->

๋ณธ ์–ธ์–ด๋ชจ๋ธ์˜ ์„ฑ๋Šฅํ‰๊ฐ€๋Š” ์—ฐ๊ตฌ ๋ณด๊ณ ์„œ ๊ณผํ•™๊ธฐ์ˆ ํ‘œ์ค€๋ถ„๋ฅ˜ ํƒœ์Šคํฌ์— ํŒŒ์ธํŠœ๋‹ํ•˜์—ฌ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜์˜€์œผ๋ฉฐ, ๊ทธ ๊ฒฐ๊ณผ๋Š” ์•„๋ž˜์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.
-  ์—ฐ๊ตฌ ๋ณด๊ณ ์„œ ๊ณผํ•™๊ธฐ์ˆ ํ‘œ์ค€๋ถ„๋ฅ˜ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์…‹(doi.org/10.23057/50), 145 Classes, 209,454 Training Set, 89,767 Test Set

#### Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->
F1-micro/macro: ์ •๋‹ต Top3 ์ค‘ ์ตœ์†Œ 1๊ฐœ ์˜ˆ์ธก์‹œ ์„ฑ๊ณต ๊ธฐ์ค€

F1-strict: ์ •๋‹ต Top3 ์ค‘ ์˜ˆ์ธกํ•œ ์ˆ˜ ๋งŒํผ ์„ฑ๊ณต ๊ธฐ์ค€

### Results

F1-micro: 0.85, F1-macro: 0.52, F1-strict: 0.71



## Technical Specifications

### Model Objective

MLM is a technique in which you take your tokenized sample and replace some of the tokens with the < mask > token and train your model with it. The model then tries to predict what should come in the place of that < mask > token and gradually starts learning about the data. MLM teaches the model about the relationship between words.

Eg. Suppose you have a sentence - 'Deep Learning is so cool! I love neural networks.', now replace few words with the < mask > token.

Masked Sentence - 'Deep Learning is so < mask >! I love < mask > networks.'

### Compute Infrastructure

KISTI ๊ตญ๊ฐ€์Šˆํผ์ปดํ“จํŒ…์„ผํ„ฐ NEURON ์‹œ์Šคํ…œ. HPE ClusterStor E1000, Lustre, Slurm

#### Hardware

NVIDIA A100 80G GPU 24EA

#### Software

Python 3.8, Cuda 10.2, PyTorch 1.10

## Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
ํ•œ๊ตญ๊ณผํ•™๊ธฐ์ˆ ์ •๋ณด์—ฐ๊ตฌ์› (2023) : ํ•œ๊ตญ์–ด ๊ณผํ•™๊ธฐ์ˆ ๋ถ„์•ผ DeBERTa ์‚ฌ์ „ํ•™์Šต ๋ชจ๋ธ (KorSciDeBERTa). Version 1.0. ํ•œ๊ตญ๊ณผํ•™๊ธฐ์ˆ ์ •๋ณด์—ฐ๊ตฌ์›.


## Model Card Authors

๊น€๊ฒฝ๋ฏผ, ๊น€์„ฑ์ฐฌ, ๊น€์€ํฌ, ์ด๋ฏผํ˜ธ, ์ด์Šน์šฐ. ํ•œ๊ตญ๊ณผํ•™๊ธฐ์ˆ ์ •๋ณด์—ฐ๊ตฌ์› ์ธ๊ณต์ง€๋Šฅ๋ฐ์ดํ„ฐ์—ฐ๊ตฌ๋‹จ

## Model Card Contact

๊น€๊ฒฝ๋ฏผ, kkmkorea kisti.re.kr