maximazzik commited on
Commit
c91ade5
·
verified ·
1 Parent(s): 7bdd2fe

docs: align dataset card with GitHub MAS repository

Browse files
Files changed (1) hide show
  1. README.md +92 -40
README.md CHANGED
@@ -1,6 +1,7 @@
1
  ---
2
  language:
3
  - ar
 
4
  task_categories:
5
  - image-to-text
6
  tags:
@@ -8,7 +9,12 @@ tags:
8
  - handwritten-text-recognition
9
  - historical-documents
10
  - arabic
 
 
 
11
  license: other
 
 
12
  configs:
13
  - config_name: mas
14
  data_files:
@@ -18,61 +24,107 @@ configs:
18
  path: mas/test-*
19
  ---
20
 
21
- # MAS: Medieval Arabic Script
22
 
23
- MAS is a line-level OCR benchmark containing 11,841 expert-annotated lines
24
- from nine authentic Arabic manuscript works. It covers Naskh, Taliq, and
25
- Nastaliq scripts and manuscripts copied between the 12th and 20th centuries.
 
26
 
27
- This dataset accompanies the ICDAR 2026 paper *A Millennium of Arabic
28
- Manuscripts in Three Styles: A Line-Level OCR Benchmark for Naskh, Taliq,
29
- and Nastaliq*.
30
 
31
- Source code, training configurations, and evaluation utilities are available
32
- in the [MAS GitHub repository](https://github.com/ai-forever/MAS).
33
 
34
- ## Dataset structure
 
 
 
35
 
36
- The dataset exposes one configuration, `mas`, with two splits:
37
 
38
- - `train`: 9,472 lines
39
- - `test`: 2,369 lines
40
 
41
- The local development aliases `uzbek_article_train` and `uzbek_article_val`
42
- are not used in this release. The local validation file is published as the
43
- `test` split, matching the paper.
 
44
 
45
- Each row contains:
 
 
46
 
47
- - `image`: the line image;
48
- - `text`: diplomatic transcription;
49
- - `prompt`: exact OCR instruction used in the experiments;
50
- - `messages`: ShareGPT-style user/assistant messages;
51
- - `script`, `domain`, `domain_label`, `work`, `author`, `author_dates`;
52
- - `composition_date`, `copy_date`, `manuscript_id`;
53
- - `source_pack` and `line_id` for traceability.
54
 
55
- Transcriptions preserve original spelling, punctuation, dots, diacritics, and
56
- archaic forms. No normalization, correction, or modernization is applied.
57
-
58
- ## Split protocol
59
-
60
- The split is 80/20 at image level with no page overlap. All nine works occur
61
- in both splits; therefore, results may overestimate generalization to entirely
62
- unseen manuscripts.
63
-
64
- ## License
65
 
66
- License terms for manuscript images and annotations must be confirmed with
67
- the owning institutions before this repository is made public.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
68
 
69
  ## Citation
70
 
 
 
71
  ```bibtex
72
- @inproceedings{novopoltsev2026mas,
73
  title={A Millennium of Arabic Manuscripts in Three Styles: A Line-Level OCR Benchmark for Naskh, Taliq, and Nastaliq},
74
- author={Novopoltsev, Maxim and Murtazin, Ruslan and Sakhovskiy, Andrey and others},
75
- booktitle={International Conference on Document Analysis and Recognition},
76
- year={2026}
 
 
 
77
  }
78
  ```
 
1
  ---
2
  language:
3
  - ar
4
+ pretty_name: MAS
5
  task_categories:
6
  - image-to-text
7
  tags:
 
9
  - handwritten-text-recognition
10
  - historical-documents
11
  - arabic
12
+ - naskh
13
+ - taliq
14
+ - nastaliq
15
  license: other
16
+ size_categories:
17
+ - 10K<n<100K
18
  configs:
19
  - config_name: mas
20
  data_files:
 
24
  path: mas/test-*
25
  ---
26
 
27
+ # MAS: A Millennium of Arabic Manuscripts in Three Styles
28
 
29
+ **MAS (Medieval Arabic Script)** is a line-level OCR benchmark of 11,841
30
+ expert-annotated lines from nine authentic Arabic manuscript works. It covers
31
+ Naskh, Taliq, and Nastaliq. Copies date from the 12th to the 20th century;
32
+ compositions from the 10th to the early 20th.
33
 
34
+ This dataset accompanies the ICDAR 2026 paper
35
+ [*A Millennium of Arabic Manuscripts in Three Styles: A Line-Level OCR Benchmark
36
+ for Naskh, Taliq, and Nastaliq*](https://link.springer.com/chapter/10.1007/978-3-032-36033-5_38).
37
 
38
+ Training configurations, the `lmms-eval` task, and evaluation notes are in the
39
+ [MAS GitHub repository](https://github.com/ai-forever/MAS).
40
 
41
+ <p align="center">
42
+ <img src="assets/calligraphy-styles.jpg" alt="Naskh, Taliq, and Nastaliq page examples from MAS" width="900">
43
+ </p>
44
+ <p align="center"><em>Calligraphic styles in MAS, left to right: Naskh, Taliq, Nastaliq.</em></p>
45
 
46
+ ## Splits
47
 
48
+ The dataset exposes one configuration, `mas`:
 
49
 
50
+ | Split | Lines | Share |
51
+ |---|---:|---:|
52
+ | `train` | 9,472 | 80% |
53
+ | `test` | 2,369 | 20% |
54
 
55
+ The split is 80/20 at image level with no page overlap. All nine works occur in
56
+ both splits, so this is not a manuscript-held-out evaluation. The published
57
+ `test` split is the held-out evaluation set used in the paper.
58
 
59
+ ```python
60
+ from datasets import load_dataset
 
 
 
 
 
61
 
62
+ dataset = load_dataset("maximazzik/MAS", "mas")
63
+ print(dataset)
64
+ print(dataset["train"][0].keys())
65
+ ```
 
 
 
 
 
 
66
 
67
+ ## Coverage
68
+
69
+ The paper groups the collection into **five domains**: astronomy, history,
70
+ mathematics, religion, and Sufi literature.
71
+
72
+ The `domain` field uses six machine-readable labels. `siyasat` (advice to rulers)
73
+ and `prayers` together correspond to the paper's Religion group:
74
+
75
+ | Paper domain | `domain` | `domain_label` |
76
+ |---|---|---|
77
+ | Astronomy | `astronomy` | Astronomy |
78
+ | History | `history` | History |
79
+ | Mathematics | `mathematics` | Mathematics |
80
+ | Religion | `siyasat` | Advice to rulers (siyasat-nama) |
81
+ | Religion | `prayers` | Prayer collection |
82
+ | Sufi literature | `sufi_lit` | Sufi literature |
83
+
84
+ ## Record schema
85
+
86
+ | Field | Meaning |
87
+ |---|---|
88
+ | `image` | Embedded line image |
89
+ | `text` | Diplomatic UTF-8 transcription |
90
+ | `prompt` | OCR instruction used to construct the sample |
91
+ | `messages` | ShareGPT-style user/assistant messages |
92
+ | `script` | Naskh, Taliq, or Nastaliq |
93
+ | `domain`, `domain_label` | Machine-readable and display domain names |
94
+ | `work`, `author`, `author_dates` | Bibliographic metadata |
95
+ | `composition_date`, `copy_date` | Work and manuscript chronology |
96
+ | `manuscript_id` | Stable identifier for one of the nine works |
97
+ | `source_pack`, `line_id` | Source grouping and line-level traceability |
98
+ | `split` | `train` or `test` |
99
+
100
+ ## Annotation
101
+
102
+ Transcriptions are diplomatic: original spelling, punctuation, dots, diacritics,
103
+ and historical forms are preserved without modernization or correction. Each
104
+ line was independently transcribed by two trained researchers, with disagreements
105
+ resolved through cross-validation and discussion.
106
+
107
+ ## Evaluation
108
+
109
+ Open-source LVLMs were evaluated with
110
+ [`lmms-eval`](https://github.com/EvolvingLMMs-Lab/lmms-eval). Specialized OCR
111
+ systems and closed-source models were scored from saved predictions using the
112
+ same CER/WER definitions. See
113
+ [`docs/reproduction.md`](https://github.com/ai-forever/MAS/blob/main/docs/reproduction.md)
114
+ and [`docs/results.md`](https://github.com/ai-forever/MAS/blob/main/docs/results.md).
115
 
116
  ## Citation
117
 
118
+ If you use the MAS dataset, please cite:
119
+
120
  ```bibtex
121
+ @inproceedings{novopoltsev2027mas,
122
  title={A Millennium of Arabic Manuscripts in Three Styles: A Line-Level OCR Benchmark for Naskh, Taliq, and Nastaliq},
123
+ author={Novopoltsev, Maxim and Murtazin, Ruslan and Sakhovskiy, Andrey and Bojarskaja, Emilia and Kokh, Vladimir and Ulitin, Ivan and Abdullayev, Botirjon and Aminov, Khamidulla and Ismoilov, Masudkhon and Budennyy, Semen},
124
+ booktitle={Document Analysis and Recognition -- ICDAR 2026},
125
+ pages={643--659},
126
+ year={2027},
127
+ publisher={Springer Nature Switzerland},
128
+ doi={10.1007/978-3-032-36033-5_38}
129
  }
130
  ```