Kayra-1 is a ~100M parameter Turkish language model.
This project does NOT aim to compete with large LLMs.
Its goal is to study how small Turkish models behave under instruction tuning,
where they fail, and how much improvement is possible with limited resources.
Current observations:
- Simple factual questions often succeed.
- Open-ended questions may hallucinate.
- Tokenization issues are visible in Turkish words.
- Reasoning is weak by design.
Kayra-1 is intentionally kept small to make iteration fast
and improvements measurable.
This model is experimental and shared openly
to document the learning process.