tascib/turkish-instruction

---
license: cc-by-sa-4.0
language:
- tr
tags:
- instruction-tuning
- turkish
- nlp
- llm
size_categories:
- 100K<n<1M
---

# Turkish Instruction Dataset

## Dataset Description

This dataset is a large-scale Turkish instruction-tuning corpus created by combining multiple publicly available datasets and applying cleaning and deduplication steps. It is designed for training and evaluating large language models (LLMs) in Turkish.

The dataset was prepared as part of a capstone project by students from Sabancı University.

## Data Sources

The dataset is constructed from the following sources:

- **TFLai / Turkish-Alpaca** (Apache-2.0)
https://huggingface.co/datasets/TFLai/Turkish-Alpaca

- **OpenAssistant / oasst1** (Apache-2.0)
https://huggingface.co/datasets/OpenAssistant/oasst1

- **atasoglu / databricks-dolly-15k-tr** (CC BY-SA 3.0)
https://huggingface.co/datasets/atasoglu/databricks-dolly-15k-tr

- **merve / turkish_instructions** (Apache-2.0)
https://huggingface.co/datasets/merve/turkish_instructions

- **umarigan / openhermes_tr** (license not explicitly specified, derived from OpenHermes)
https://huggingface.co/datasets/umarigan/openhermes_tr/tree/main

## Dataset Statistics

- **Total collected samples:** 360,351
- **After deduplication:** 324,080
- **Duplicates removed:** 36,271

### Source Breakdown

- alpaca-turkish: 51,862
- oasst1-turkish: 9
- dolly-turkish: 14,714
- openhermes-turkish: 225,867
- merve-turkish: 31,628

## Preprocessing

- Merging multiple datasets
- Cleaning inconsistent or malformed samples
- Removing duplicates

## Intended Use

This dataset is intended primarily for:

- Supervised fine-tuning (SFT) of instruction-following language models
- Training Turkish instruction-tuned LLMs
- NLP research and experimentation in Turkish

This dataset is not intended for pretraining from scratch, as it consists of instruction-response pairs rather than raw text corpora.

## Data Format

Each sample follows an instruction-tuning format:

- instruction: Task description given to the model
- input: Optional additional context
- output: Expected response

This structure is compatible with common instruction-tuning pipelines such as Alpaca, Dolly, and OpenAI-style SFT datasets.

## License

This dataset is released under the **CC BY-SA 4.0** license.

Note:
- Some source datasets are licensed under Apache-2.0 and CC BY-SA 3.0.
- Derived datasets must comply with the most restrictive applicable terms.
- Users must provide proper attribution and share derivative works under the same license.

## Limitations

- May contain noise due to automatic aggregation
- Potential biases inherited from source datasets
- Uneven distribution across sources (OpenHermes dominates)

## Acknowledgements

We thank the creators of the original datasets:

- TFLai
- OpenAssistant
- Databricks / Dolly
- Merve Noyan
- OpenHermes contributors

## Disclaimer

No responsibility or liability is accepted for the use of this dataset.