QuantaSparkLabs commited on
Commit
eb0502c
ยท
verified ยท
1 Parent(s): f6b3731

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +182 -0
README.md ADDED
@@ -0,0 +1,182 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ - es
5
+ - fr
6
+ - de
7
+ - it
8
+ - pt
9
+ - nl
10
+ - ru
11
+ - zh
12
+ - ja
13
+ - ko
14
+ - ar
15
+ - hi
16
+ - tr
17
+ - vi
18
+
19
+ license: apache-2.0
20
+
21
+ pipeline_tag: translation
22
+
23
+ tags:
24
+ - translation
25
+ - multilingual
26
+ - t5
27
+ - opus100
28
+ - whirlwindai
29
+ ---
30
+
31
+ <p align="center">
32
+
33
+ <img src="https://capsule-render.vercel.app/api?type=venom&height=230&text=&animation=fadeIn&color=gradient&customColorList=12,20,24,30"/>
34
+
35
+ <img src="https://readme-typing-svg.demolab.com?font=Space+Grotesk&weight=700&size=28&pause=2500&color=00E7FF&center=true&vCenter=true&width=760&lines=Translate-15L;15+Languages.;Powered+by+WhirlwindAI"/>
36
+
37
+ </p>
38
+
39
+ <img src="https://img.shields.io/badge/Parameters-60M-4FC3F7?style=for-the-badge">
40
+ <img src="https://img.shields.io/badge/Languages-15-8B5CF6?style=for-the-badge">
41
+ <img src="https://img.shields.io/badge/Framework-Transformers-06B6D4?style=for-the-badge">
42
+ <img src="https://img.shields.io/badge/License-Apache--2.0-10B981?style=for-the-badge">
43
+
44
+ </div>
45
+
46
+ ---
47
+
48
+ # Overview
49
+
50
+ Translate-15L is a lightweight multilingual translation model trained on OPUS100.
51
+
52
+ Rather than maximizing parameter count, the model focuses on providing practical multilingual translation with fast inference and a compact footprint.
53
+
54
+ It supports translation between **English** and **14 additional languages** in both directions.
55
+
56
+ ---
57
+
58
+ # Quick Start
59
+
60
+ ```python
61
+ from transformers import AutoTokenizer, T5ForConditionalGeneration
62
+
63
+ model = T5ForConditionalGeneration.from_pretrained(
64
+ "WhirlwindAI/Translate-15L"
65
+ )
66
+
67
+ tokenizer = AutoTokenizer.from_pretrained(
68
+ "WhirlwindAI/Translate-15L"
69
+ )
70
+
71
+ text = "Hello, how are you?"
72
+
73
+ prompt = f"translate en to fr: {text}"
74
+
75
+ inputs = tokenizer(prompt, return_tensors="pt")
76
+
77
+ outputs = model.generate(
78
+ **inputs,
79
+ max_new_tokens=50,
80
+ num_beams=4
81
+ )
82
+
83
+ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
84
+ ```
85
+
86
+ ---
87
+
88
+ # Supported Languages
89
+
90
+ | Code | Language |
91
+ |------|----------|
92
+ | ๐Ÿ‡ฌ๐Ÿ‡ง en | English |
93
+ | ๐Ÿ‡ช๐Ÿ‡ธ es | Spanish |
94
+ | ๐Ÿ‡ซ๐Ÿ‡ท fr | French |
95
+ | ๐Ÿ‡ฉ๐Ÿ‡ช de | German |
96
+ | ๐Ÿ‡ฎ๐Ÿ‡น it | Italian |
97
+ | ๐Ÿ‡ต๐Ÿ‡น pt | Portuguese |
98
+ | ๐Ÿ‡ณ๐Ÿ‡ฑ nl | Dutch |
99
+ | ๐Ÿ‡ท๐Ÿ‡บ ru | Russian |
100
+ | ๐Ÿ‡จ๐Ÿ‡ณ zh | Chinese |
101
+ | ๐Ÿ‡ฏ๐Ÿ‡ต ja | Japanese |
102
+ | ๐Ÿ‡ฐ๐Ÿ‡ท ko | Korean |
103
+ | ๐Ÿ‡ธ๐Ÿ‡ฆ ar | Arabic |
104
+ | ๐Ÿ‡ฎ๐Ÿ‡ณ hi | Hindi |
105
+ | ๐Ÿ‡น๐Ÿ‡ท tr | Turkish |
106
+ | ๐Ÿ‡ป๐Ÿ‡ณ vi | Vietnamese |
107
+
108
+ ---
109
+
110
+ # Performance
111
+
112
+ <div align="center">
113
+
114
+ ## BLEU Evaluation
115
+
116
+ <img src="bleu_separate.png" width="96%">
117
+
118
+ <br><br>
119
+
120
+ ## Language Overview
121
+
122
+ <img src="radar_chart.png" width="70%">
123
+
124
+ </div>
125
+
126
+ ---
127
+
128
+ # Highlights
129
+
130
+ | Direction | BLEU |
131
+ |-----------|------:|
132
+ | ๐Ÿ‡ฌ๐Ÿ‡ง โ†’ ๐Ÿ‡ซ๐Ÿ‡ท | **32.43** |
133
+ | ๐Ÿ‡ฉ๐Ÿ‡ช โ†’ ๐Ÿ‡ฌ๐Ÿ‡ง | **16.93** |
134
+ | ๐Ÿ‡ฌ๐Ÿ‡ง โ†’ ๐Ÿ‡ช๐Ÿ‡ธ | 6.51 |
135
+ | ๐Ÿ‡ซ๐Ÿ‡ท โ†’ ๐Ÿ‡ฌ๐Ÿ‡ง | 5.13 |
136
+ | ๐Ÿ‡ฌ๐Ÿ‡ง โ†’ ๐Ÿ‡ต๐Ÿ‡น | 5.05 |
137
+ | ๐Ÿ‡ฌ๐Ÿ‡ง โ†’ ๐Ÿ‡ฎ๐Ÿ‡น | 4.15 |
138
+
139
+ ---
140
+
141
+ # Speed
142
+
143
+ <div align="center">
144
+
145
+ # โšก ~8,151 Tokens / Second
146
+
147
+ Fast enough for lightweight multilingual applications while remaining compact.
148
+
149
+ See **speed.txt** for the complete benchmark.
150
+
151
+ </div>
152
+
153
+ ---
154
+
155
+ # Examples
156
+
157
+ <div align="center">
158
+
159
+ <img src="download.png" width="48%">
160
+ <img src="download (1).png" width="48%">
161
+
162
+ </div>
163
+
164
+ ---
165
+
166
+ # Notes
167
+
168
+ - English โ†” 14 Languages
169
+ - Optimized for compact deployment
170
+ - Trained on OPUS100
171
+ - Best performance on high-resource European languages
172
+ - Performance on low-resource languages remains limited
173
+
174
+ ---
175
+
176
+ <div align="center">
177
+
178
+ ### Built by WhirlwindAI
179
+
180
+ Efficient AI โ€ข Practical Research โ€ข Open Models
181
+
182
+ </div>