README.md

---
dataset_info:
  features:
  - name: prompt
    dtype: string
  - name: answer
    dtype: string
  - name: labels
    list:
    - name: end
      dtype: int64
    - name: label
      dtype: string
    - name: start
      dtype: int64
  - name: split
    dtype: string
  - name: task_type
    dtype: string
  - name: dataset
    dtype: string
  - name: language
    dtype: string
  splits:
  - name: train
    num_bytes: 66155103
    num_examples: 17790
  download_size: 17097014
  dataset_size: 66155103
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: mit
task_categories:
- question-answering
- text-generation
- summarization
- text-classification
- text-retrieval
language:
- tr
tags:
- turkish
- hallucination-detection
- rag
- low-resource
- multilingual
---

# RAGTruth-TR

`newmindai/RAGTruth-TR` is a Turkish-translated version of the [`wandb/RAGTruth-processed`](https://huggingface.co/datasets/wandb/RAGTruth-processed) dataset.
It is designed for evaluating **Retrieval-Augmented Generation (RAG)** systems in Turkish, enabling research in hallucination detection, fact-checking, and response quality assessment.

---

## Dataset Summary

* **Source Dataset:** [`wandb/RAGTruth-processed`](https://huggingface.co/datasets/wandb/RAGTruth-processed)
* **Target Language:** Turkish
* **Purpose:** Hallucination detection and RAG evaluation in Turkish NLP systems
* **License:** MIT (inherits from the original dataset)

This dataset follows the same schema as the original RAGTruth-processed dataset but provides high-quality Turkish translations.

---

## Dataset Creation

The dataset was originally derived from the **RAGTruth** dataset, which contains responses from retrieval-augmented generation models annotated for hallucinations.

For this version:

* The English dataset was **translated into Turkish**.
* Translation was performed using **Gemma 3 27B**.
* We used a **single NVIDIA L40S GPU** and served the model through **vLLM** for efficient translation.

This ensures **consistent terminology** and **domain-specific fidelity** in the Turkish version.

---

## Languages

* **Turkish (tr)**
* Original dataset: **English (en)**

---

## Dataset Structure

The structure mirrors the original `wandb/RAGTruth-processed`.

Each example includes:

* `id`: Unique identifier
* `question`: The user query (translated to Turkish)
* `context`: Retrieved passages (translated to Turkish)
* `answer`: Generated response (translated to Turkish)
* `hallucination_label`: Annotation for hallucinations (inherited, unchanged)

---

## Example

```json
{
  "id": "12345",
  "question": "Türkçeye çevrilmiş bir örnek soru",
  "context": [
    "Türkçeye çevrilmiş bir bağlam pasajı."
  ],
  "answer": "Türkçeye çevrilmiş model cevabı.",
  "hallucination_label": "no_hallucination"
}
```

---

## Intended Uses

* Evaluate **hallucination detection** methods in Turkish.
* Fine-tune or evaluate **RAG models** in Turkish.
* Use as a **benchmark dataset** for multilingual hallucination detection.

---

## Citation

If you use this dataset, please cite both the original and this translated version, and our paper: 

```bibtex
@misc{ragtruth-tr,
  author       = {newmindai},
  title        = {RAGTruth-TR: Turkish Translation of RAGTruth},
  year         = {2025},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/newmindai/RAGTruth-TR}
}

@misc{ragtruth-processed,
  author       = {Weights & Biases},
  title        = {RAGTruth-processed},
  year         = {2025},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/wandb/RAGTruth-processed}
}

@article{turklettuceDetect2025,
  title={Turk-LettuceDetect: A Hallucination Detection Models for Turkish RAG Applications},
  author={Selva Taş, Mahmut El Huseyni, Özay Ezerceli, Reyhan Bayraktar, Fatma Betül Terzioğlu},
  journal={arXiv preprint arXiv:2509.17671},
  year={2025}
}

```