File size: 34,019 Bytes
ef0c646
 
 
 
 
c7c64db
ef0c646
c7c64db
ef0c646
 
c7c64db
ef0c646
c7c64db
ef0c646
42497fb
 
ef0c646
42497fb
 
ef0c646
42497fb
 
ef0c646
42497fb
 
 
 
 
 
 
 
ef0c646
c7c64db
 
 
ef0c646
c7c64db
ef0c646
c7c64db
 
 
 
ef0c646
c7c64db
ef0c646
c7c64db
 
 
42497fb
 
ef0c646
c7c64db
ef0c646
c7c64db
 
 
 
ef0c646
c7c64db
ef0c646
42497fb
ef0c646
42497fb
 
 
 
ef0c646
c7c64db
 
 
 
42497fb
 
ef0c646
42497fb
ef0c646
42497fb
 
ef0c646
42497fb
ef0c646
42497fb
 
 
 
ef0c646
42497fb
 
 
 
 
 
 
 
 
 
c7c64db
 
42497fb
 
 
ef0c646
42497fb
ef0c646
42497fb
 
c7c64db
 
 
 
 
 
 
 
 
 
42497fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ef0c646
c7c64db
ef0c646
42497fb
ef0c646
42497fb
 
 
ef0c646
c7c64db
ef0c646
42497fb
 
c7c64db
ef0c646
42497fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ef0c646
42497fb
ef0c646
42497fb
ef0c646
42497fb
ef0c646
42497fb
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
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
---
library_name: transformers
tags: []
---

이 토크나이저는 [microsoft/Phi-4-mini-instruct](https://huggingface.co/microsoft/Phi-4-mini-instruct)에 [devngho/jamo-tokenizer-exp1](https://huggingface.co/devngho/jamo-tokenizer-exp1)를 병합해 만들어졌습니다.

NFKD 정규화를 하면 한국어 외 다른 언어 토크나이징에 영향을 줄 수 있습니다. 따라서 NFKD 정규화를 토크나이저에 적용하는 대신 미리 NFKD 정규화를 한글 글자에만 적용하고 Replace하도록 했습니다.


**한국어 예시**: 원래 토크나이저보다 효율적으로 토크나이징

Text: `모든 인간은 태어날 때부터 자유로우며 그 존엄과 권리에 있어 동등하다. 인간은 천부적으로 이성과 양심을 부여받았으며 서로 형제애의 정신으로 행동하여야 한다.`

phi4: 85 (0.00% more effective than phi-4)
�|�|든| 인|간|은| �|�|어|�|�| 때|부|터| 자|�|�|로|우|�|�| 그| �|��|�|�|과| �|�|�|리|에| 있|어| �|�|�|�|하|다|.| 인|간|은| �|�|부|적|으로| 이|성|과| �|�|�|�|을| 부|여|�|�|�|�|으|�|�| 서|로| �|�|제|�|�|의| 정|신|으로| �|��|동|하여|야| 한|다|.

phi4-jamo: 39 (117.95% more effective than phi-4)
모든| 인간은| 태어나|ᆯ 때|부터| 자유|로우|며| 그| 존|엄|과| 권리|에 있어| 동|등|하다|.| 인간은| 천|부|적으로| 이|성과| 양|심을| 부여|받아|ᆻ으며| 서로| 형제|애|의 정|신|으로| 행동|하여|야 한다|.

phi4-mini: 51 (66.67% more effective than phi-4)
모|든| 인간|은| 태|어|날| 때|부터| 자유|로|우|며| 그| 존|엄|과| 권|리에| 있어| 동|등|하다|.| 인간|은| 천|부|적으로| 이|성과| 양|심|을| 부|여|받|았|으며| 서로| 형|제|애|의| 정신|으로| 행동|하여|야| 한다|.

phi4-mini-jamo: 39 (117.95% more effective than phi-4)
모든| 인간은| 태어나|ᆯ 때|부터| 자유|로우|며| 그| 존|엄|과| 권리|에 있어| 동|등|하다|.| 인간은| 천|부|적으로| 이|성과| 양|심을| 부여|받아|ᆻ으며| 서로| 형제|애|의 정|신|으로| 행동|하여|야 한다|.

jamo-exp1: 39 (117.95% more effective than phi-4)
모든|인간은|태어나|ᆯ 때|부터|자유|로우|며|그|존|엄|과|권리|에 있어|동|등|하다|.|인간은|천|부|적으로|이|성과|양|심을|부여|받아|ᆻ으며|서로|형제|애|의 정|신|으로|행동|하여|야 한다|.

gpt-4o: 51 (66.67% more effective than phi-4)
모|든| 인간|은| 태|어|날| 때|부터| 자유|로|우|며| 그| 존|엄|과| 권|리에| 있어| 동|등|하다|.| 인간|은| 천|부|적으로| 이|성과| 양|심|을| 부|여|받|았|으며| 서로| 형|제|애|의| 정신|으로| 행동|하여|야| 한다|.

Text:
```
PyTorch, TensorFlow, JAX를 위한 최첨단 머신러닝

🤗 Transformers는 사전학습된 최첨단 모델들을 쉽게 다운로드하고 훈련시킬 수 있는 API와 도구를 제공합니다. 사전학습된 모델을 쓰면 컴퓨팅 비용과 탄소 배출량이 줄고, 모델을 처음부터 훈련시키는 데 필요한 시간과 리소스를 절약할 수 있습니다. 저희 모델들은 다양한 분야의 태스크를 지원합니다.

📝 자연어 처리: 텍스트 분류, 개체명 인식, 질의응답, 언어 모델링, 요약, 번역, 객관식 질의응답, 텍스트 생성
🖼️ 컴퓨터 비전: 이미지 분류, 객체 탐지, 객체 분할
🗣️ 오디오: 자동음성인식, 오디오 분류
🐙 멀티모달: 표 질의응답, 광학 문자 인식 (OCR), 스캔한 문서에서 정보 추출, 비디오 분류, 시각 질의응답

🤗 Transformers는 PyTorch, TensorFlow와 JAX 간의 상호운용성을 지원합니다. 유연하게 모델의 각 단계마다 다른 프레임워크를 사용할 수도 있습니다. 예를 들어 코드 3줄만 써서 모델을 훈련시킨 다음, 다른 프레임워크 상에서 추론할 수 있습니다. 모델을 운영 환경에 배포하기 위해 ONNX나 TorchScript 형식으로 내보낼 수도 있습니다.

커뮤니티에 참여하시려면 Hub, 포럼, 디스코드를 방문해주세요!
```
<details> <summary>토큰 단위 결과</summary>
phi4: 540 (0.00% more effective than phi-4)
Py|T|orch|,| TensorFlow|,| J|AX|를| 위|한| 최|�|�|�|�| �|�|�|신|러|�|�|

|�|�|�| Transformers|는| 사|전|�|�|습|된| 최|�|�|�|�| 모|�|�|들|을| �|�|�|게| 다|운|로|드|하|고| �|�|�|�|�|시|�|��| 수| 있는| API|와| �|�|구|를| 제|공|합니다|.| 사|전|�|�|습|된| 모|�|�|을| �|�|�|면| �|�|�|�|�|�|�| 비|용|과| �|�|소| 배|출|�|�|이| �|�|고|,| 모|�|�|을| 처|음|부|터| �|�|�|�|�|시|�|��|는| �|��| 필|요|한| 시|간|과| 리|소|스|를| �|�|�|�|할| 수| 있|습니다|.| �|�|�|�|�| 모|�|�|들|은| 다|�|�|한| �|�|야|의| �|�|스|크|를| 지|원|합니다|.

|�|�|�| 자|�|�|어| 처리|:| �|��|스트| �|�|류|,| 개|체|명| 인|식|,| �|�|의|�|�|�|�|,| �|�|어| 모|�|�|�|�|,| 요|�|�|,| 번|�|�|,| �|�|�|��|식| �|�|의|�|�|�|�|,| �|��|스트| 생성|
|�|�|�|️| �|�|�|�|�|터| 비|전|:| 이미|지| �|�|류|,| 객체| �|�|지|,| 객체| �|�|할|
|�|�|�|️| 오|디|오|:| 자|동|음|성|인|식|,| 오|디|오| �|�|류|
|�|�|�| �|�|�|�|�|�|�|�|�|:| �|��| �|�|의|�|�|�|�|,| �|�|�|�|�| 문자| 인|식| (|OCR|),| �|�|�|�|�|한| 문|서|에서| 정보| 추|출|,| 비|디|오| �|�|류|,| 시|�|�| �|�|의|�|�|�|�|

|�|�|�| Transformers|는| Py|T|orch|,| TensorFlow|와| J|AX| �|�|의| 상|호|운|용|성|을| 지|원|합니다|.| �|�|�|�|하|게| 모|�|�|의| �|�| �|�|�|�|�|�다| 다|른| 프|�|�|임|�|�|크|를| 사용|할| 수|도| 있|습니다|.| 예|를| �|�|어| 코드| |3|�|�|만| �|�|�|서| 모|�|�|을| �|�|�|�|�|시|�|��| 다|음|,| 다|른| 프|�|�|임|�|�|크| 상|에서| 추|�|�|할| 수| 있|습니다|.| 모|�|�|을| �|�|�|�| �|�|경|에| 배|�|�|�|하기| 위|해| ON|NX|나| Torch|Script| �|�|식|으로| 내|보|�|�| 수|도| 있|습니다|.

|�|�|�|�|�|�|니|�|�|�|에| �|�|여|하|시|�|�|면| Hub|,| 포|�|�|,| �|�|스|�|��드|를| 방|문|해|주세요|!

phi4-jamo: 291 (85.57% more effective than phi-4)
Py|T|orch|,| TensorFlow|,| J|AX|를| 위한| 최|첨|단| 머|신|러|닝|
|
|�|�|�| Transformers|는| 사전|학습|된| 최|첨|단| 모델|들을| 쉽게| 다|운|로드|하고| 훈련|시키|ᆯ 수 있는| API|와| 도|구를| 제공합니다|.| 사전|학습|된| 모델을| 쓰|면| ᄏ|ᅥᆷ퓨|팅| 비용|과| 탄소| 배출|량이| 줄|고|,| 모델을| 처음|부터| 훈련|시키는 데| 필요한| 시간과| 리|소|스를| 절|약하|ᆯ 수 있습니다|.| 저|희| 모델|들은| 다양한| 분야|의| 태|스크|를| 지원|합니다|.

|�|�|�| 자연|어| 처리|:| 텍|스트| 분류|,| 개|체|명| 인식|,| 질의|응답|,| 언어| 모데|ᆯ링|,| 요약|,| 번|역|,| 객|관|식| 질의|응답|,| 텍|스트| 생성|
|�|�|�|️| 컴퓨터| 비|전|:| 이미지| 분류|,| 객|체| 탐|지|,| 객|체| 분|할|
|�|�|�|️| 오|디오|:| 자동|음성|인식|,| 오|디오| 분류|
|�|�|�| 멀티|모|달|:| 표| 질의|응답|,| 광학| 문자| 인식| |(|OCR|),| 스|캐|ᆫ|한| 문|서|에서| 정보| 추출|,| 비디오| 분류|,| 시각| 질의|응답|
|
|�|�|�| Transformers|는| Py|T|orch|,| TensorFlow|와| J|AX| 간|의 상호|운용|성을| 지원|합니다|.| 유연|하게| 모데|ᆯ의| 각| 단계|마다| 다른| 프레임|워크|를 사용|할 수도| 있습니다|.| 예를 들어| 코드| |3|줄|만| 써|서| 모델을| 훈련|시킨| 다음|,| 다른| 프레임|워크| 상에서| 추|론|할 수 있|습니다|.| 모델을| 운영| 환경|에| 배|포|하기 위해| ON|NX|나| Torch|Script| 형|식으로| 내|보내|ᆯ 수도| 있습니다|.

|커뮤니티|에 참여|하시|려면| Hub|,| 포|럼|,| 디|스코|드를| 방문|해주|세요|!

phi4-mini: 353 (52.97% more effective than phi-4)
Py|Torch|,| Tensor|Flow|,| J|AX|를| 위한| 최|첨|단| 머신|러|닝|

|�|�| Transformers|는| 사|전|학|습|된| 최|첨|단| 모델|들을| 쉽게| 다운로드|하고| �|��|련|시|킬| 수| 있는| API|와| 도|구|를| 제공합니다|.| 사|전|학|습|된| 모델|을| 쓰|면| 컴|퓨|팅| 비용|과| 탄|소| 배|출|량|이| 줄|고|,| 모델|을| 처음|부터| �|��|련|시키|는| 데| 필요한| 시간|과| 리|소|스를| 절|약|할| 수| 있습니다|.| 저|희| 모델|들은| 다양한| 분야|의| 태|스크|를| 지원|합니다|.

|�|��| 자연|어| 처리|:| �|�|스트| 분|류|,| 개|체|명| 인|식|,| 질|의|응|답|,| 언|어| 모델|링|,| 요|약|,| 번|역|,| 객|관|식| 질|의|응|답|,| �|�|스트| 생성|
|�|�|�|️| 컴|퓨터| 비|전|:| 이미지| 분|류|,| 객체| 탐|지|,| 객체| 분|할|
|�|�|�|️| 오|디오|:| 자동|음|성|인|식|,| 오|디오| 분|류|
|�|�| �|�|티|모|달|:| 표| 질|의|응|답|,| 광|학| 문자| 인|식| (|OCR|),| 스|�|�|한| 문|서|에서| 정보| 추|출|,| 비|디오| 분|류|,| 시|각| 질|의|응|답|

|�|�| Transformers|는| Py|Torch|,| Tensor|Flow|와| J|AX| 간|의| 상|호|운|용|성을| 지원|합니다|.| 유|연|하게| 모델|의| 각| 단계|마다| 다른| 프|레|임|워크|를| 사용할| 수도| 있습니다|.| 예|를| 들어| 코드| |3|줄|만| �|�|서| 모델|을| �|��|련|시|킨| 다음|,| 다른| 프|레|임|워크| 상|에서| 추|론|할| 수| 있습니다|.| 모델|을| 운영| 환경|에| 배|포|하기| 위해| ON|NX|나| Torch|Script| 형|식|으로| 내|보|낼| 수도| 있습니다|.

|커|뮤니|티|에| 참여|하시|려|면| Hub|,| 포|럼|,| 디|스|코|드를| 방문|해주세요|!

phi4-mini-jamo: 287 (88.15% more effective than phi-4)
Py|Torch|,| Tensor|Flow|,| J|AX|를| 위한| 최|첨|단| 머|신|러|닝|
|
|�|�| Transformers|는| 사전|학습|된| 최|첨|단| 모델|들을| 쉽게| 다|운|로드|하고| 훈련|시키|ᆯ 수 있는| API|와| 도|구를| 제공합니다|.| 사전|학습|된| 모델을| 쓰|면| ᄏ|ᅥᆷ퓨|팅| 비용|과| 탄소| 배출|량이| 줄|고|,| 모델을| 처음|부터| 훈련|시키는 데| 필요한| 시간과| 리|소|스를| 절|약하|ᆯ 수 있습니다|.| 저|희| 모델|들은| 다양한| 분야|의| 태|스크|를| 지원|합니다|.

|�|��| 자연|어| 처리|:| 텍|스트| 분류|,| 개|체|명| 인식|,| 질의|응답|,| 언어| 모데|ᆯ링|,| 요약|,| 번|역|,| 객|관|식| 질의|응답|,| 텍|스트| 생성|
|�|�|�|️| 컴퓨터| 비|전|:| 이미지| 분류|,| 객|체| 탐|지|,| 객|체| 분|할|
|�|�|�|️| 오|디오|:| 자동|음성|인식|,| 오|디오| 분류|
|�|�| 멀티|모|달|:| 표| 질의|응답|,| 광학| 문자| 인식| |(O|CR|),| 스|캐|ᆫ|한| 문|서|에서| 정보| 추출|,| 비디오| 분류|,| 시각| 질의|응답|
|
|�|�| Transformers|는| Py|Torch|,| Tensor|Flow|와| J|AX| 간|의 상호|운용|성을| 지원|합니다|.| 유연|하게| 모데|ᆯ의| 각| 단계|마다| 다른| 프레임|워크|를 사용|할 수도| 있습니다|.| 예를 들어| 코드| |3|줄|만| 써|서| 모델을| 훈련|시킨| 다음|,| 다른| 프레임|워크| 상에서| 추|론|할 수 있|습니다|.| 모델을| 운영| 환경|에| 배|포|하기 위해| ON|NX|나| Torch|Script| 형|식으로| 내|보내|ᆯ 수도| 있습니다|.

|커뮤니티|에 참여|하시|려면| Hub|,| 포|럼|,| 디|스코|드를| 방문|해주|세요|!

jamo-exp1: 345 (56.52% more effective than phi-4)
P|y|T|or|ch|,|T|en|s|or|F|lo|w|,|J|A|X|를 위한|최|첨|단|머|신|러|닝|
|
|�|�|�|�|T|r|an|s|for|m|ers|는 사|전|학습|된|최|첨|단|모델|들을|쉽게|다|운|로드|하고|훈련|시키|ᆯ 수 있는|AP|I|와|도|구를|제공합니다|.|사전|학습|된|모델을|쓰|면|ᄏ|ᅥᆷ퓨|팅|비용|과|탄소|배출|량이|줄|고|,|모델을|처음|부터|훈련|시키는 데|필요한|시간과|리|소|스를|절|약하|ᆯ 수 있습니다|.|저|희|모델|들은|다양한|분야|의|태|스크|를|지원|합니다|.|
|
|�|�|�|�|자연|어|처리|:|텍|스트|분류|,|개|체|명|인식|,|질의|응답|,|언어|모데|ᆯ링|,|요약|,|번|역|,|객|관|식|질의|응답|,|텍|스트|생성|
|�|�|�|�|�|�|�|컴퓨터|비|전|:|이미지|분류|,|객|체|탐|지|,|객|체|분|할|
|�|�|�|�|�|�|�|오|디오|:|자동|음성|인식|,|오|디오|분류|
|�|�|�|�|멀티|모|달|:|표|질의|응답|,|광학|문자|인식||(|OC|R|)|,|스|캐|ᆫ|한|문|서|에서|정보|추출|,|비디오|분류|,|시각|질의|응답|
|
|�|�|�|�|T|r|an|s|for|m|ers|는|P|y|T|or|ch|,|T|en|s|or|F|lo|w|와|J|A|X|간|의 상호|운용|성을|지원|합니다|.|유연|하게|모데|ᆯ의|각|단계|마다|다른|프레임|워크|를 사용|할 수도|있습니다|.|예를 들어|코드||3|줄|만|써|서|모델을|훈련|시킨|다음|,|다른|프레임|워크|상에서|추|론|할 수 있|습니다|.|모델을|운영|환경|에|배|포|하기 위해|ON|N|X|나|T|or|ch|S|c|ri|p|t|형|식으로|내|보내|ᆯ 수도|있습니다|.|
|
|커뮤니티|에 참여|하시|려면|H|ub|,|포|럼|,|디|스코|드를|방문|해주|세요|!

gpt-4o: 353 (52.97% more effective than phi-4)
Py|Torch|,| Tensor|Flow|,| J|AX|를| 위한| 최|첨|단| 머신|러|닝|

|�|�| Transformers|는| 사|전|학|습|된| 최|첨|단| 모델|들을| 쉽게| 다운로드|하고| �|��|련|시|킬| 수| 있는| API|와| 도|구|를| 제공합니다|.| 사|전|학|습|된| 모델|을| 쓰|면| 컴|퓨|팅| 비용|과| 탄|소| 배|출|량|이| 줄|고|,| 모델|을| 처음|부터| �|��|련|시키|는| 데| 필요한| 시간|과| 리|소|스를| 절|약|할| 수| 있습니다|.| 저|희| 모델|들은| 다양한| 분야|의| 태|스크|를| 지원|합니다|.

|�|��| 자연|어| 처리|:| �|�|스트| 분|류|,| 개|체|명| 인|식|,| 질|의|응|답|,| 언|어| 모델|링|,| 요|약|,| 번|역|,| 객|관|식| 질|의|응|답|,| �|�|스트| 생성|
|�|�|�|️| 컴|퓨터| 비|전|:| 이미지| 분|류|,| 객체| 탐|지|,| 객체| 분|할|
|�|�|�|️| 오|디오|:| 자동|음|성|인|식|,| 오|디오| 분|류|
|�|�| �|�|티|모|달|:| 표| 질|의|응|답|,| 광|학| 문자| 인|식| (|OCR|),| 스|�|�|한| 문|서|에서| 정보| 추|출|,| 비|디오| 분|류|,| 시|각| 질|의|응|답|

|�|�| Transformers|는| Py|Torch|,| Tensor|Flow|와| J|AX| 간|의| 상|호|운|용|성을| 지원|합니다|.| 유|연|하게| 모델|의| 각| 단계|마다| 다른| 프|레|임|워크|를| 사용할| 수도| 있습니다|.| 예|를| 들어| 코드| |3|줄|만| �|�|서| 모델|을| �|��|련|시|킨| 다음|,| 다른| 프|레|임|워크| 상|에서| 추|론|할| 수| 있습니다|.| 모델|을| 운영| 환경|에| 배|포|하기| 위해| ON|NX|나| Torch|Script| 형|식|으로| 내|보|낼| 수도| 있습니다|.

|커|뮤니|티|에| 참여|하시|려|면| Hub|,| 포|럼|,| 디|스|코|드를| 방문|해주세요|!</details>

phi4: 540 (0.00% more effective than phi-4)

phi4-jamo: 291 (85.57% more effective than phi-4)

phi4-mini: 353 (52.97% more effective than phi-4)

phi4-mini-jamo: 287 (88.15% more effective than phi-4)

jamo-exp1: 345 (56.52% more effective than phi-4)

gpt-4o: 353 (52.97% more effective than phi-4)

**다국어 예시**: 원래 토크나이저와 동일

Text:
```
State-of-the-art Machine Learning for PyTorch, TensorFlow, and JAX.

🤗 Transformers provides APIs and tools to easily download and train state-of-the-art pretrained models. Using pretrained models can reduce your compute costs, carbon footprint, and save you the time and resources required to train a model from scratch. These models support common tasks in different modalities, such as:

📝 Natural Language Processing: text classification, named entity recognition, question answering, language modeling, code generation, summarization, translation, multiple choice, and text generation.
🖼️ Computer Vision: image classification, object detection, and segmentation.
🗣️ Audio: automatic speech recognition and audio classification.
🐙 Multimodal: table question answering, optical character recognition, information extraction from scanned documents, video classification, and visual question answering.

🤗 Transformers support framework interoperability between PyTorch, TensorFlow, and JAX. This provides the flexibility to use a different framework at each stage of a model’s life; train a model in three lines of code in one framework, and load it for inference in another. Models can also be exported to a format like ONNX and TorchScript for deployment in production environments.

Join the growing community on the Hub, forum, or Discord today!
```

<details> <summary>토큰 단위 결과</summary>
phi4: 261 (0.00% more effective than phi-4)
State|-of|-the|-art| Machine| Learning| for| Py|T|orch|,| TensorFlow|,| and| J|AX|.

|�|�|�| Transformers| provides| APIs| and| tools| to| easily| download| and| train| state|-of|-the|-art| pretrained| models|.| Using| pretrained| models| can| reduce| your| compute| costs|,| carbon| footprint|,| and| save| you| the| time| and| resources| required| to| train| a| model| from| scratch|.| These| models| support| common| tasks| in| different| modal|ities|,| such| as|:

|�|�|�| Natural| Language| Processing|:| text| classification|,| named| entity| recognition|,| question| answering|,| language| modeling|,| code| generation|,| summar|ization|,| translation|,| multiple| choice|,| and| text| generation|.
|�|�|�|️| Computer| Vision|:| image| classification|,| object| detection|,| and| segmentation|.
|�|�|�|️| Audio|:| automatic| speech| recognition| and| audio| classification|.
|�|�|�| Mult|im|odal|:| table| question| answering|,| optical| character| recognition|,| information| extraction| from| scanned| documents|,| video| classification|,| and| visual| question| answering|.

|�|�|�| Transformers| support| framework| interoper|ability| between| Py|T|orch|,| TensorFlow|,| and| J|AX|.| This| provides| the| flexibility| to| use| a| different| framework| at| each| stage| of| a| model|’s| life|;| train| a| model| in| three| lines| of| code| in| one| framework|,| and| load| it| for| inference| in| another|.| Models| can| also| be| exported| to| a| format| like| ON|NX| and| Torch|Script| for| deployment| in| production| environments|.

|Join| the| growing| community| on| the| Hub|,| forum|,| or| Discord| today|!

phi4-jamo: 261 (0.00% more effective than phi-4)
State|-of|-the|-art| Machine| Learning| for| Py|T|orch|,| TensorFlow|,| and| J|AX|.

|�|�|�| Transformers| provides| APIs| and| tools| to| easily| download| and| train| state|-of|-the|-art| pretrained| models|.| Using| pretrained| models| can| reduce| your| compute| costs|,| carbon| footprint|,| and| save| you| the| time| and| resources| required| to| train| a| model| from| scratch|.| These| models| support| common| tasks| in| different| modal|ities|,| such| as|:

|�|�|�| Natural| Language| Processing|:| text| classification|,| named| entity| recognition|,| question| answering|,| language| modeling|,| code| generation|,| summar|ization|,| translation|,| multiple| choice|,| and| text| generation|.
|�|�|�|️| Computer| Vision|:| image| classification|,| object| detection|,| and| segmentation|.
|�|�|�|️| Audio|:| automatic| speech| recognition| and| audio| classification|.
|�|�|�| Mult|im|odal|:| table| question| answering|,| optical| character| recognition|,| information| extraction| from| scanned| documents|,| video| classification|,| and| visual| question| answering|.

|�|�|�| Transformers| support| framework| interoper|ability| between| Py|T|orch|,| TensorFlow|,| and| J|AX|.| This| provides| the| flexibility| to| use| a| different| framework| at| each| stage| of| a| model|’s| life|;| train| a| model| in| three| lines| of| code| in| one| framework|,| and| load| it| for| inference| in| another|.| Models| can| also| be| exported| to| a| format| like| ON|NX| and| Torch|Script| for| deployment| in| production| environments|.

|Join| the| growing| community| on| the| Hub|,| forum|,| or| Discord| today|!

phi4-mini: 255 (2.35% more effective than phi-4)
State|-of|-the|-art| Machine| Learning| for| Py|Torch|,| Tensor|Flow|,| and| J|AX|.

|�|�| Transformers| provides| APIs| and| tools| to| easily| download| and| train| state|-of|-the|-art| pretrained| models|.| Using| pretrained| models| can| reduce| your| compute| costs|,| carbon| footprint|,| and| save| you| the| time| and| resources| required| to| train| a| model| from| scratch|.| These| models| support| common| tasks| in| different| modalities|,| such| as|:

|�|��| Natural| Language| Processing|:| text| classification|,| named| entity| recognition|,| question| answering|,| language| modeling|,| code| generation|,| summar|ization|,| translation|,| multiple| choice|,| and| text| generation|.
|�|�|�|️| Computer| Vision|:| image| classification|,| object| detection|,| and| segmentation|.
|�|�|�|️| Audio|:| automatic| speech| recognition| and| audio| classification|.
|�|�| Mult|im|odal|:| table| question| answering|,| optical| character| recognition|,| information| extraction| from| scanned| documents|,| video| classification|,| and| visual| question| answering|.

|�|�| Transformers| support| framework| interoperability| between| Py|Torch|,| Tensor|Flow|,| and| J|AX|.| This| provides| the| flexibility| to| use| a| different| framework| at| each| stage| of| a| model|’s| life|;| train| a| model| in| three| lines| of| code| in| one| framework|,| and| load| it| for| inference| in| another|.| Models| can| also| be| exported| to| a| format| like| ON|NX| and| Torch|Script| for| deployment| in| production| environments|.

|Join| the| growing| community| on| the| Hub|,| forum|,| or| Discord| today|!

phi4-mini-jamo: 255 (2.35% more effective than phi-4)
State|-of|-the|-art| Machine| Learning| for| Py|Torch|,| Tensor|Flow|,| and| J|AX|.

|�|�| Transformers| provides| APIs| and| tools| to| easily| download| and| train| state|-of|-the|-art| pretrained| models|.| Using| pretrained| models| can| reduce| your| compute| costs|,| carbon| footprint|,| and| save| you| the| time| and| resources| required| to| train| a| model| from| scratch|.| These| models| support| common| tasks| in| different| modalities|,| such| as|:

|�|��| Natural| Language| Processing|:| text| classification|,| named| entity| recognition|,| question| answering|,| language| modeling|,| code| generation|,| summar|ization|,| translation|,| multiple| choice|,| and| text| generation|.
|�|�|�|️| Computer| Vision|:| image| classification|,| object| detection|,| and| segmentation|.
|�|�|�|️| Audio|:| automatic| speech| recognition| and| audio| classification|.
|�|�| Mult|im|odal|:| table| question| answering|,| optical| character| recognition|,| information| extraction| from| scanned| documents|,| video| classification|,| and| visual| question| answering|.

|�|�| Transformers| support| framework| interoperability| between| Py|Torch|,| Tensor|Flow|,| and| J|AX|.| This| provides| the| flexibility| to| use| a| different| framework| at| each| stage| of| a| model|’s| life|;| train| a| model| in| three| lines| of| code| in| one| framework|,| and| load| it| for| inference| in| another|.| Models| can| also| be| exported| to| a| format| like| ON|NX| and| Torch|Script| for| deployment| in| production| environments|.

|Join| the| growing| community| on| the| Hub|,| forum|,| or| Discord| today|!

jamo-exp1: 710 (-63.24% more effective than phi-4)
St|ate|-|of|-|the|-|art|M|ac|hi|n|e|L|e|ar|n|ing|for|P|y|T|or|ch|,|T|en|s|or|F|lo|w|,|and|J|A|X|.|
|
|�|�|�|�|T|r|an|s|for|m|ers|pro|v|id|es|AP|I|s|and|to|ol|s|to|e|a|si|ly|d|ow|n|lo|ad|and|tr|ain|st|ate|-|of|-|the|-|art|p|r|et|r|ain|ed|m|od|el|s|.|U|si|ng|p|r|et|r|ain|ed|m|od|el|s|c|an|r|ed|uc|e|y|our|com|p|ut|e|c|o|st|s|,|c|ar|b|on|f|o|ot|p|r|in|t|,|and|s|av|e|y|ou|the|t|im|e|and|r|es|our|c|es|re|qu|ir|ed|to|tr|ain|a|m|od|el|f|ro|m|s|c|r|at|ch|.|T|h|es|e|m|od|el|s|s|up|p|ort|com|m|on|t|as|k|s|in|d|if|f|er|ent|m|od|al|it|ies|,|s|u|ch|as|:|
|
|�|�|�|�|N|at|ur|al|L|ang|u|ag|e|P|ro|c|es|si|ng|:|t|ex|t|c|l|as|si|f|ic|ation|,|n|am|ed|ent|ity|r|ec|og|n|it|ion|,|qu|est|ion|an|s|w|er|ing|,|l|ang|u|ag|e|m|od|el|ing|,|c|od|e|g|en|er|ation|,|s|um|m|ar|iz|ation|,|tr|an|s|l|ation|,|m|ult|ip|le|ch|o|ic|e|,|and|t|ex|t|g|en|er|ation|.|
|�|�|�|�|�|�|�|C|om|p|ut|er|V|i|sion|:|im|ag|e|c|l|as|si|f|ic|ation|,|ob|j|ect|d|et|ect|ion|,|and|s|eg|ment|ation|.|
|�|�|�|�|�|�|�|A|ud|i|o|:|a|ut|om|at|ic|sp|e|ec|h|r|ec|og|n|it|ion|and|a|ud|i|o|c|l|as|si|f|ic|ation|.|
|�|�|�|�|M|ult|im|od|al|:|t|ab|le|qu|est|ion|an|s|w|er|ing|,|op|t|ical|ch|ar|ac|ter|r|ec|og|n|it|ion|,|in|for|m|ation|ex|tr|act|ion|f|ro|m|s|c|an|n|ed|d|oc|um|ent|s|,|v|id|e|o|c|l|as|si|f|ic|ation|,|and|v|is|u|al|qu|est|ion|an|s|w|er|ing|.|
|
|�|�|�|�|T|r|an|s|for|m|ers|s|up|p|ort|f|r|am|ew|or|k|in|ter|op|er|ab|il|ity|b|et|w|e|en|P|y|T|or|ch|,|T|en|s|or|F|lo|w|,|and|J|A|X|.|T|hi|s|pro|v|id|es|the|f|le|x|ib|il|ity|to|us|e|a|d|if|f|er|ent|f|r|am|ew|or|k|at|e|ac|h|st|ag|e of|a|m|od|el|�|�|�|s|l|if|e|;|tr|ain|a|m|od|el|in|th|re|e|l|in|es|of|c|od|e|in|on|e|f|r|am|ew|or|k|,|and|lo|ad|it|for|in|f|er|enc|e|in|an|ot|h|er|.|M|od|el|s|c|an|al|s|o|b|e|ex|p|ort|ed|to|a|for|m|at|l|i|k|e|ON|N|X|and|T|or|ch|S|c|ri|p|t|for|d|ep|lo|y|ment|in|pro|d|uc|tion|en|v|ir|on|ment|s|.|
|
|J|o|in|the|g|ro|w|ing|com|m|un|ity|on|the|H|ub|,|for|um|,|or|D|is|c|or|d|t|od|ay|!

gpt-4o: 255 (2.35% more effective than phi-4)
State|-of|-the|-art| Machine| Learning| for| Py|Torch|,| Tensor|Flow|,| and| J|AX|.

|�|�| Transformers| provides| APIs| and| tools| to| easily| download| and| train| state|-of|-the|-art| pretrained| models|.| Using| pretrained| models| can| reduce| your| compute| costs|,| carbon| footprint|,| and| save| you| the| time| and| resources| required| to| train| a| model| from| scratch|.| These| models| support| common| tasks| in| different| modalities|,| such| as|:

|�|��| Natural| Language| Processing|:| text| classification|,| named| entity| recognition|,| question| answering|,| language| modeling|,| code| generation|,| summar|ization|,| translation|,| multiple| choice|,| and| text| generation|.
|�|�|�|️| Computer| Vision|:| image| classification|,| object| detection|,| and| segmentation|.
|�|�|�|️| Audio|:| automatic| speech| recognition| and| audio| classification|.
|�|�| Mult|im|odal|:| table| question| answering|,| optical| character| recognition|,| information| extraction| from| scanned| documents|,| video| classification|,| and| visual| question| answering|.

|�|�| Transformers| support| framework| interoperability| between| Py|Torch|,| Tensor|Flow|,| and| J|AX|.| This| provides| the| flexibility| to| use| a| different| framework| at| each| stage| of| a| model|’s| life|;| train| a| model| in| three| lines| of| code| in| one| framework|,| and| load| it| for| inference| in| another|.| Models| can| also| be| exported| to| a| format| like| ON|NX| and| Torch|Script| for| deployment| in| production| environments|.

|Join| the| growing| community| on| the| Hub|,| forum|,| or| Discord| today|!
</details>

phi4: 261 (0.00% more effective than phi-4)

phi4-jamo: 261 (0.00% more effective than phi-4)

phi4-mini: 255 (2.35% more effective than phi-4)

phi4-mini-jamo: 255 (2.35% more effective than phi-4)

jamo-exp1: 710 (-63.24% more effective than phi-4)

gpt-4o: 255 (2.35% more effective than phi-4)