DeeDe-2B
Experimental Turkish-first Large Language Model
DeeDe-2B is an experimental 2.4B parameter language model developed as an independent research project. The primary goal of this project is to explore whether a capable Turkish language model can be built with limited computational resources while establishing a foundation for future iterations.
This is not a production-ready model. It is released to encourage community feedback and to accelerate future improvements.
Overview
DeeDe-2B is designed around a Turkish-first tokenizer instead of relying on multilingual tokenization. The project focuses on improving Turkish language efficiency and serving as a research platform for future DeeDe AI models.
Current limitations include:
- Limited factual/world knowledge
- Weak reasoning capability
- Difficulty handling complex multi-step tasks
- Noticeable repetition in some responses
- Not suitable for production use
Despite these limitations, the project demonstrates that a small independent team—or even a single developer—can train a Turkish language model with modest resources.
Model Details
| Attribute | Value |
|---|
| Model Name | DeeDe-2B |
| Parameters | 2.4 Billion |
| Architecture | Decoder-only Transformer |
| Tokenizer | Kumru's Tokenizer |
| Training Data | 140 GB real text |
| Tokens Processed | ~66 Billion |
| Fine-tuning | Fully Synthetic SFT Dataset |
| Status | Experimental |
Motivation
High-quality Turkish conversational datasets are extremely limited.
Most publicly available datasets are either:
- machine translated,
- low quality,
- repetitive,
- or unsuitable for instruction tuning.
Because of this, much of the instruction tuning relied on synthetic data generation.
Synthetic data works surprisingly well in early stages, but eventually causes the model to become repetitive and less helpful.
The long-term goal is to gradually replace synthetic examples with carefully curated human preference data.
Intended Use
DeeDe-2B is intended for:
- Turkish NLP research
- Language model experimentation
- Prompt engineering research
- Tokenizer evaluation
- Educational purposes
It is not recommended for:
- Medical advice
- Legal advice
- Financial decisions
- Safety-critical applications
- Production deployments
Current Limitations
The model may:
- hallucinate facts,
- generate incorrect information,
- misunderstand complicated instructions,
- produce repetitive outputs,
- fail at advanced reasoning tasks.
Please verify important information independently.
Future Roadmap
The project aims to improve through several iterations:
- Better Turkish instruction datasets
- Human preference data
- Community feedback integration
- Larger-scale continued pretraining
- Improved reasoning
- Better factual accuracy
- More diverse response generation
Ultimately, DeeDe AI is planned to become the intelligence layer behind the DeeDe Search Engine and other future DeeDe products.
Community Feedback
Community feedback is one of the most valuable resources for improving the model.
If you encounter poor responses, please consider sharing:
- the prompt,
- the model's answer,
- and the response you expected instead.
These examples will help build higher-quality supervised fine-tuning datasets for future versions.
Training Philosophy
This project is intentionally open about its current limitations.
Rather than claiming benchmark-leading performance, the objective is to openly document the process of building a Turkish language model from limited resources and continuously improving it with community participation.
Every version should be measurably better than the previous one.
Acknowledgements
Thank you to everyone who tests the model, reports issues, shares ideas, or contributes data and infrastructure.
Building capable open Turkish language models is a long-term effort, and every contribution helps move the project forward.
Citation
1@misc{deede2b,
2 title={DeeDe-2B},
3 author={Uğurhan Çolak},
4 year={2026},
5 publisher={Hugging Face},
6 note={Experimental Turkish-first Large Language Model}
7}
DeeDe AI
Building AI for Turkish, one iteration at a time.
