Spaces:

---

Efe2898

/

Gemma

like 0

Running

Hi Hugging Face team and everyone,

I’m working on a Gemma 3 1B model focused on Turkish reasoning. I use merged and distilled datasets (from models like GPT-OSS and Qwen).

SLM's have a huge problem: Even 3B–7B models are not good at reasoning in Turkish. They understand English well, but Turkish outputs feel unnatural and sometimes wrong.

My firstly goal is to build a small 1B model that can do strong and natural reasoning in Turkish. I want it to perform like bigger models, but better in Turkish.

This project is open-source if i can. I will share:

• training process (CPT → SFT → reasoning)
• dataset building method
• notebooks (Kaggle / TPU)

I’m applying for a GPU grant to:

• let people test the model easily
• get real feedback from the community
• improve the model with that feedback

I believe small models can be very strong if trained correctly, especially for non-English languages.

Thanks for your time and support.