A very critical development has occurred in Türkiye in terms of developing local and national artificial intelligence solutions, reducing foreign dependency and strengthening the open source ecosystem. Within the scope of open source studies carried out within Trendyol Group, the large-scale SFT (Supervised Fine-Tuning) information set “Turkish Atlas”, designed for the training of Turkish artificial intelligence models, was officially shared.

This data set, which contains nearly 336 thousand structured examples in total, aims to enable language models to understand Turkish commands more accurately and naturally. The project, which is purified from subjective content and distilled entirely from new open-access web resources, brings a new breath to artificial intelligence integration processes in Turkey.

• Large Scale Dataset:Turkish Atlas contains nearly 336,000 structured information examples specially optimized for training native artificial intelligence models.

• Multitasking Support:dataset; It contains critical tasks that improve the basic capabilities of language models, such as command tracking, summarization, question-answering and producing structured output.

• Open Source and Ready Format:The project, which was presented to the open source ecosystem with the support of Trendyol Group, was prepared in the “system-user-assistant” speech structure and JSONL format so that it can be directly integrated into SFT training processes.

What is Turkish Atlas and Why is it Important for Artificial Intelligence?

Large language models (LLM) need to go through controlled fine-tuning (PFT) processes to master a reasonable language and accurately carry out human commands. Turkish Atlas steps in at this point, aiming to fill the gap in high-quality information that local developers and researchers need the most.

This information set, created by meticulously scanning current open access web resources, is produced not by uploading raw data directly to the system, but by filtering artificial intelligence models in the most appropriate way for the learning process. By cleaning the subjective, biased or unqualified texts it contains, it becomes easier for the models to be trained to give more reliable and balanced outputs.

The Format That Changes the Standards in Artificial Intelligence Education

As a technical infrastructure, Turkish Atlas is structured to be directly compatible with contemporary artificial intelligence architectures. To make it easier for developers, all recordings are prepared by adhering to the standard chat template familiar to today’s popular language models:

• System:It determines the role the model must assume and the rules it must follow.

• User:It contains the question or command that the human poses to the model.

• Assistant:It symbolizes the most accurate and optimized response that the model should give.

Since this hierarchical structure is presented in JSONL (JSON Lines) format, data scientists can directly include the information in the training cycle without any preliminary process or conversion effort.

Capabilities of Turkish Language Models Increase

The 336 thousand examples offered by the project focus on developing all the basic muscles that language models will need in daily and professional work processes, rather than focusing on a single area. The four basic tasks included in the data set directly determine the functionality of domestic artificial intelligence solutions.

Command Tracking and Q&A Performance

The success of a language model depends on understanding exactly what the user wants. Turkish Atlas, with its structure containing complex command sequences, ensures that the models fulfill the instructions completely. In addition, question-answer examples fed from a wide knowledge pool increase the information consistency of the models and their ability to produce natural responses that comply with the rules of Turkish language knowledge.

Summarization and Structured Output Generation

Particularly in corporate business processes, analysis and reporting of long texts is of great importance. Summarization tasks in the dataset help the models produce refined information without missing the main idea in the texts. The ability to produce structured output means that models are not just plain text; It allows you to present data in tables, code blocks or certain formats. This significantly facilitates the integration of artificial intelligence into companies’ existing software infrastructures.

Open Source Ecosystem and the Goal of Reducing Foreign Dependency

In the technology world, removing artificial intelligence technologies from the monopoly of global actors has a strategic importance in terms of countries asserting their own data sovereignty. Comprehensive and qualified data set projects such as Turkish Atlas lay the groundwork for the local artificial intelligence ecosystem in Turkey to meet global standards.

This work, which came to life with Trendyol Group’s strong support to the open source ecosystem, is available to everyone, from personal developers to large technology companies. In this way, inefficient formulas for training huge models with billions of parameters from scratch or trying to train artificial intelligence by translating foreign language data into Turkish are prevented. This data set, distilled entirely from domestic resources, signals that the domestic artificial intelligence assistants, customer service bots and corporate information analysis tools we will see in the future will be much more capable.