shyngys879/Kazakh-Wiki-RAG-Dataset

Kazakh Wiki RAG Dataset

Kazakh Wiki RAG Dataset is a Kazakh-language query-passage dataset built from Kazakh Wikipedia text for semantic search, retrieval-augmented generation, and embedding model training.

The dataset was created to support Kazakh-language information retrieval and low-resource language embedding research.

Dataset Format

The main dataset uses a query-passage retrieval format.

Main columns:

• anchor — Kazakh retrieval query
• positive — relevant Kazakh passage
• source — source label

Example row structure:

anchor:   Жерлеу рәсімін жасауға мүмкіндік болмаған жағдайлар
positive: Жерлеу рәсімдерін жасауға мүмкіндік болмайтын түрлі апат жағдайларынан...
source:   wiki_llm

Intended Uses

This dataset can be used for:

• semantic search
• retrieval-augmented generation
• embedding model training
• query-passage matching
• Kazakh retrieval evaluation
• hard-negative retrieval experiments
• low-resource language retrieval research

Main Files

• train.csv — training split
• dev.csv — development split
• test.csv — test split
• train_pairs.csv, dev_pairs.csv, test_pairs.csv — pair-format retrieval data
• train_triplets.csv — triplet-format training data with hard negatives
• eval_pairs.json — evaluation pairs
• quality_audit.csv — quality audit examples
• manual_audit_sample.csv — manual audit sample
• best_accepted_samples.csv — high-quality accepted examples
• worst_accepted_samples.csv — lower-confidence accepted examples
• suspicious_samples.csv — suspicious examples identified during filtering
• diagnostics_report.txt — dataset diagnostics
• production_readiness_report.txt — production-readiness summary
• production_30k_readiness_report.txt — 30K dataset readiness report
• dataset_stats.json — dataset statistics
• dataset_card_stats.json — dataset-card statistics

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset(
    "shyngys879/Kazakh-Wiki-RAG-Dataset",
    data_files={
        "train": "train.csv",
        "dev": "dev.csv",
        "test": "test.csv",
    }
)

print(dataset)
print(dataset["train"][0])

Sentence Transformers Usage

The dataset can be used for contrastive embedding training with Sentence Transformers.

from datasets import load_dataset
from sentence_transformers import InputExample

dataset = load_dataset(
    "shyngys879/Kazakh-Wiki-RAG-Dataset",
    data_files={"train": "train.csv"}
)

train_examples = [
    InputExample(texts=[row["anchor"], row["positive"]])
    for row in dataset["train"]
]

Data Source

The dataset is based on Kazakh Wikipedia-derived passages. Retrieval-style queries were generated and filtered for query-passage training.

Quality Control

The dataset generation process included:

• query-passage relevance filtering
• suspicious sample detection
• duplicate and leakage checks
• source distribution analysis
• query type distribution analysis
• hard negative diagnostics
• manual audit samples
• production-readiness reporting

Limitations

This dataset may contain noise because some queries were generated automatically. It should be used for research, experimentation, and retrieval model development, not as a source of guaranteed factual truth.

Since the underlying passages are based on Wikipedia-derived content, the dataset may inherit factual gaps, outdated information, or topical imbalance from Wikipedia.

License

Because the dataset is based on Wikipedia-derived content, it is released under CC BY-SA 4.0.

Related Model

This dataset was used to support the development of Kazakh retrieval embedding models.

• Model: https://huggingface.co/shyngys879/kazakh-e5-rag-embedding

Citation

If you use this dataset, please cite this Hugging Face repository.