File size: 9,049 Bytes
b0099cf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb19af7
b0099cf
 
 
 
 
 
 
 
eb19af7
b0099cf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c3c59f6
 
 
b0099cf
c3c59f6
 
 
 
b0099cf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c3c59f6
 
 
b0099cf
c3c59f6
b0099cf
 
c3c59f6
 
b0099cf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: mit
task_categories:
- text-generation
- feature-extraction
- text-classification
language:
- hi
- en
tags:
- hindiwikipedia
- hiwiki
- hindi
- english
- noisy-to-clean
- text-cleaning
- aadimind
- ADITYA
pretty_name: THE SONA CORPUS
size_categories:
- 100K<n<1M
---

# THE SONA CORPUS — Noisy-to-Clean Hindi–English Parallel Dataset

A clean, bilingual dataset card you can read at a glance and use immediately.

- Curated by: Aditya (AADIMIND)
- Languages: Hindi, English
- Total examples: 581312 (INPUT: 256 TOKEN• TARGET: 256 TOKEN)
- Tasks: Text cleaning, GEC, OCR post-processing, Seq2Seq fine-tuning
- License: MIT
- Source: Hindi Wikipedia (HiWiki) processed into noisy–clean pairs
- Repo: https://huggingface.co/datasets/AADIMIND/sona-corpus

---

## Overview

THE SONA CORPUS is a parallel dataset of noisy inputs and their cleaned/corrected targets in Hindi and English. It’s built for models that transform imperfect text (typos, OCR errors, informal writing) into clean, standardized text.

- Input: Noisy text (with errors)
- Target: Corrected / clean text
- Great for: Seq2Seq training, GEC, OCR post-processing, text normalization

---

## Dataset Structure

- Columns:
  - input_text: string (noisy text)
  - target_text: string (clean/corrected text)

- SIZE OF DATA:
  - TRAIN LINES OF DATASET: 581312

  - INPUT: 256 TOKEN

  - TARGET: 256 TOKEN

---

## Data Sources and Creation

- Source data: Hindi Wikipedia dump (HiWiki)
- Processing:
  - Extracted sentences from HiWiki
  - Added artificial noise and OCR-like errors
  - Aligned with manually cleaned target text
- Annotations: Expert-generated (Aditya); manual cleaning + selective automated scripts

---

## Intended Uses

- Direct:
  - Train encoder–decoder (Seq2Seq) models for Hindi text correction
  - Benchmark for noisy-to-clean conversion
  - OCR post-processing and spell correction
- Out-of-scope:
  - Conversational chatbot training
  - Sensitive/private data processing

---

## Limitations and Notes

- Domain: Primarily Wikipedia-style text; limited colloquial/code-mixed coverage
- Generalization: May require adaptation for social-media style, heavy code-mixing, or domain-specific jargon
- License note: If source content includes Wikipedia text (CC BY-SA), ensure compliance with attribution and share-alike requirements when redistributing derivatives

---

## How to Use

```python
from datasets import load_dataset

dataset = load_dataset("AADIMIND/sona-corpus")

# 95% train और 5% validation split
dataset = dataset["train"].train_test_split(test_size=0.05)

train_data = dataset["train"]
val_data = dataset["test"]

print(train_data[0])
print(val_data[0])

```

Model training tip:
- For encoder–decoder models (e.g., T5, mT5), concatenate language/task prompts if helpful (e.g., "gec: <text>").
- Evaluate with normalization-aware metrics (e.g., chrF, token-level accuracy, or edit distance).

---

## Citation

BibTeX:
```bibtex
@misc{aadimind-sona-corpus-2025,
  title={THE SONA CORPUS: Noisy-to-Clean Hindi-English Text Dataset},
  author={Aditya},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/AADIMIND/sona-corpus}}
}
```

APA:
Aditya. (2025). THE SONA CORPUS: Noisy-to-Clean Hindi-English Text Dataset. Hugging Face. https://huggingface.co/datasets/AADIMIND/sona-corpus

---

## Maintainer

- Aditya (AADIMIND)
- Contact: https://huggingface.co/AADIMIND
- EMAIL : AK332575@GMAIL.COM
- INSTAGRAM : @thinker.aditya
---

# THE SONA CORPUS — डेटासेट कार्ड (हिंदी)

- क्यूरेटर: आदित्य (AADIMIND)
- भाषाएँ: हिंदी, अंग्रेज़ी
- कुल उदाहरण: 38,497 (ट्रेन: 36,572 • वैलिडेशन: 1,925)
- कार्य: टेक्स्ट क्लीनिंग, GEC, OCR पोस्ट-प्रोसेसिंग, Seq2Seq फाइन-ट्यूनिंग
- लाइसेंस: MIT
- स्रोत: हिंदी विकिपीडिया (HiWiki) से प्रोसेस्ड नॉइज़ी–क्लीन पेयर्स
- रिपो: https://huggingface.co/datasets/AADIMIND/sona-corpus

---

## परिचय

THE SONA CORPUS में नॉइज़ी इनपुट और उनके साफ/सही (करेक्टेड) टारगेट पेयर्स हैं। इसका उद्देश्य ऐसे मॉडल ट्रेन करना है जो गलत/अधूरा टेक्स्ट (टाइपो, OCR त्रुटियाँ, अनौपचारिक लेखन) को मानकीकृत, साफ टेक्स्ट में बदल सकें।

- इनपुट: नॉइज़ी टेक्स्ट (त्रुटियों सहित)
- टारगेट: साफ/सही किया हुआ टेक्स्ट
- उपयोग: Seq2Seq प्रशिक्षण, GEC, OCR पोस्ट-प्रोसेसिंग, टेक्स्ट नॉर्मलाइज़ेशन

---

## डेटासेट संरचना

- कॉलम:
  - input_text: string (नॉइज़ी टेक्स्ट)
  - target_text: string (क्लीन/करेक्टेड टेक्स्ट)

- स्प्लिट्स:
  - ट्रेन: 36,572
  - वैलिडेशन: 1,925
  - कुल: 38,497

---

## स्रोत और निर्माण

- स्रोत डेटा: हिंदी विकिपीडिया (HiWiki)
- प्रोसेस:
  - HiWiki से वाक्य निकाले गए
  - कृत्रिम नॉइज़ व OCR-जैसी त्रुटियाँ जोड़ी गईं
  - मैन्युअल रूप से साफ/करेक्टेड टारगेट के साथ एलाइन किया गया
- एनोटेशन: विशेषज्ञ (आदित्य) द्वारा—मैन्युअल क्लीनिंग + चयनित ऑटो स्क्रिप्ट्स

---

## उपयोग

- सीधे उपयोग:
  - हिंदी टेक्स्ट करेक्शन के लिए एन्कोडर–डिकोडर (Seq2Seq) मॉडल ट्रेनिंग
  - नॉइज़ी-टू-क्लीन कन्वर्ज़न के लिए बेंचमार्क
  - OCR पोस्ट-प्रोसेसिंग और स्पेल करेक्शन
- अनुपयुक्त:
  - बातचीत-आधारित चैटबॉट ट्रेनिंग
  - संवेदनशील/निजी डेटा प्रोसेसिंग

---

## सीमाएँ और नोट्स

- डोमेन: मुख्यतः विकिपीडिया-शैली; बोलचाल/कोड-मिक्स्ड कंटेंट का सीमित प्रतिनिधित्व
- सामान्यीकरण: सोशल मीडिया-स्टाइल, भारी कोड-मिक्सिंग या विशिष्ट डोमेन जार्गन के लिए अतिरिक्त अनुकूलन अपेक्षित
- लाइसेंस नोट: यदि स्रोत में विकिपीडिया टेक्स्ट (CC BY-SA) शामिल है/से व्युत्पन्न है, तो पुनर्वितरण में एट्रिब्यूशन और शेयर-अलाइक शर्तों का पालन सुनिश्चित करें

---

## उपयोग कैसे करें

```python
from datasets import load_dataset

dataset = load_dataset("AADIMIND/sona-corpus")

# 95% train और 5% validation split
dataset = dataset["train"].train_test_split(test_size=0.05)

train_data = dataset["train"]
val_data = dataset["test"]

print(train_data[0])
print(val_data[0])

```

ट्रेनिंग सुझाव:
- T5/mT5 जैसे मॉडलों के लिए टास्क/भाषा प्रॉम्प्ट जोड़ना लाभकारी हो सकता है (जैसे, "gec: <text>").
- मूल्यांकन में chrF, टोकन-स्तरीय एक्यूरेसी या एडिट-डिस्टेंस जैसे मेट्रिक उपयोगी हैं।

---

## उद्धरण

BibTeX:
```bibtex
@misc{aadimind-sona-corpus-2025,
  title={THE SONA CORPUS: Noisy-to-Clean Hindi-English Text Dataset},
  author={Aditya},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/AADIMIND/sona-corpus}}
}
```

APA:
आदित्य. (2025). THE SONA CORPUS: Noisy-to-Clean Hindi-English Text Dataset. Hugging Face. https://huggingface.co/datasets/AADIMIND/sona-corpus

---