DevHunterAI commited on
Commit
08ef48c
·
verified ·
1 Parent(s): 65df465

Add dataset card

Browse files
Files changed (1) hide show
  1. README.md +104 -0
README.md ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - tr
4
+ license: cc-by-4.0
5
+ size_categories:
6
+ - 100M<n<1B
7
+ task_categories:
8
+ - text-generation
9
+ - question-answering
10
+ tags:
11
+ - reasoning
12
+ - logic
13
+ - turkish
14
+ - pretraining
15
+ - synthetic
16
+ pretty_name: Turkish Reasoning Data
17
+ ---
18
+
19
+ # Turkish Reasoning Data
20
+
21
+ A large-scale **synthetic Turkish logic-puzzle dataset** designed for pretraining and
22
+ fine-tuning language models on structured reasoning tasks.
23
+
24
+ ## Dataset Summary
25
+
26
+ | Property | Value |
27
+ |---|---|
28
+ | Language | Turkish (`tr`) |
29
+ | Format | Plain text (pretraining) |
30
+ | Approximate tokens | ~100 million |
31
+ | Number of examples | ~720,000 |
32
+ | File size | ~450 MB |
33
+ | License | CC BY 4.0 |
34
+
35
+ ## Reasoning Categories
36
+
37
+ Each example belongs to one of eight logic-puzzle categories:
38
+
39
+ | Category (TR) | Category (EN) | Weight |
40
+ |---|---|---|
41
+ | Sıralama bulmacaları | Ordering puzzles | 25 % |
42
+ | Doğru/Yanlış ifade | True/False statement evaluation | 15 % |
43
+ | Çelişki tespiti | Contradiction detection | 15 % |
44
+ | Kim ne yaptı | Person–action–object matching | 15 % |
45
+ | Ev/kat bulmacaları | Zebra-style house/floor puzzles | 10 % |
46
+ | Seçim argümanı | Argument strength evaluation | 10 % |
47
+ | Sayısal sıralama | Numerical magnitude reasoning | 10 % |
48
+ | Zaman çizelgesi | Timeline / before–after reasoning | 10 % |
49
+
50
+ ## Format
51
+
52
+ Every example is wrapped in `<doc>...</doc>` tags and contains:
53
+
54
+ 1. **Puzzle description** — a natural-language logic problem in Turkish
55
+ 2. **Step-by-step solution** — an explicit reasoning trace leading to the answer
56
+
57
+ ```
58
+ <doc>
59
+ Aşağıdaki ipuçlarına göre Ahmet, Zeynep, Mert kişilerinin
60
+ sıralamayı belirleyin:
61
+
62
+ - Zeynep, Mert'den önce gelir.
63
+ - Ahmet 1. sıradadır.
64
+
65
+ Doğru sıralama nedir?
66
+
67
+ Adım adım çözüm:
68
+ 1. Ahmet
69
+ 2. Zeynep
70
+ 3. Mert
71
+
72
+ Sonuç: Doğru sıralama Ahmet > Zeynep > Mert şeklindedir.
73
+ </doc>
74
+ ```
75
+
76
+ ## Intended Use
77
+
78
+ - **Pretraining** Turkish language models on structured reasoning
79
+ - **Fine-tuning** (convert to SFT format by splitting on the solution separator)
80
+ - **Evaluation** of step-by-step reasoning in Turkish
81
+
82
+ ## Generation
83
+
84
+ Generated programmatically with randomized templates covering names, professions,
85
+ cities, colors, objects, and events drawn from Turkish vocabulary.
86
+ Reproducible with `random.seed(42)`.
87
+
88
+ ## Limitations
89
+
90
+ - Fully synthetic — does not cover real-world, open-ended, or commonsense reasoning
91
+ - Puzzle difficulty is limited; puzzles involve 3–5 entities
92
+ - Solutions are template-generated and may occasionally be trivially solvable
93
+
94
+ ## Citation
95
+
96
+ ```bibtex
97
+ @dataset{devhunterai2024turkish_reasoning,
98
+ title = {Turkish Reasoning Data},
99
+ author = {DevHunterAI},
100
+ year = {2025},
101
+ publisher = {Hugging Face},
102
+ url = {https://huggingface.co/datasets/DevHunterAI/turkish-reasoning-data}
103
+ }
104
+ ```