alphazhan/sauatai-ertegiler-kz-misspellings-kk-s170-len60-n6-m1-v1

SauatAI — Kazakh Misspelled Sentences from Ertegiler.kz

SauatAI is a grammar-focused dataset built from 170 children’s stories scraped from ertegiler.kz on July 5, 2025. The dataset was designed to support Kazakh language grammar correction, error detection, and text augmentation research.

📌 Dataset Details

• s170 — 170 unique stories were scraped and sentence-tokenized.
• len60 — Only sentences with ≤60 characters were retained.
• n6 — Each correct sentence has 5 corresponding corrupted variants + 1 clean version.
• m1 — Each corrupted sentence contains exactly 1 mistake:
  • The mistake is either a letter substitution (from a predefined spelling swap map) or a character omission.
  • The type is chosen randomly per sentence.
  • Ensures consistency and control over the number of errors per sample.

The final dataset consists of three CSV files:

• train.csv
• val.csv
• test.csv

Each file contains sentence-level pairs of grammatically correct and corrupted Kazakh text.

This version includes a total of 38,099 sentence pairs — 5 corrupted variants + 1 correct sentence per source (n6 setting).

🧪 Dataset Splits

Split	Count	File
Train	25,399	train.csv
Validation	6,350	val.csv
Test	6,350	test.csv

📄 Column Descriptions

Column Name	Description
story_id	Unique ID of the source story
story_title	Title of the story the sentence came from
sentence_id	Index of the sentence within the story
correct_sentence_len	Length (character count) of the correct sentence
correct_sentence	The grammatically correct version of the sentence
corrupted_sentence	A noisy version with potential spelling or character-level errors
label	0 if corrupted, 1 if correct

🧠 Intended Use

• Grammar correction model pretraining/fine-tuning
• Corrupted-to-correct sequence generation
• Evaluation of large language models on noisy Kazakh text
• Robustness testing and error correction research

📜 License

This dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to use, share, and adapt the dataset — just provide attribution.