WiroAI Turkish LLM 9B

Overview

WiroAI Turkish LLM 9B is Wiro's Turkish fine-tune of Google's Gemma 2 9B, a decoder-only transformer. It was trained with supervised fine-tuning on more than 500,000 selected Turkish instructions, using LoRA without quantization. The training targets fluent Turkish and an understanding of local context, idioms and culture, and the model also answers in other common languages. The model card lists Abdullah Bezir, Furkan Burhan Türkay and Cengiz Asmazoğlu as its authors.

On the model card, WiroAI reports a 58.0 average across the Turkish versions of MMLU, TruthfulQA, ARC, HellaSwag, GSM8K and WinoGrande, against 54.6 for Google's gemma-2-9b-it in the same evaluation setup. The widest gap is on MMLU TR: 59.8 against 51.8.

What you can build

• Turkish customer support and FAQ assistants
• Writing assistants for Turkish social posts, product descriptions and emails (the card's own example is a social media post about Istanbul)
• Summaries of Turkish articles, reports and meeting notes
• Question answering on Turkish geography, history and culture
• Rewriting Turkish text for a different tone or audience

Inputs

• Prompt (required): the user message, in Turkish or another language.
• System prompt: instructions that set the assistant's role and tone. The default is a general English assistant prompt; replace it with one written for your product.
• user_id and session_id (optional): identify a user and a conversation so chat history is kept between calls. Give each conversation its own session_id to keep histories apart.
• Sampling: temperature (0.7 by default), top_p (0.95), top_k, a repetition penalty, a switch to turn sampling off, semicolon-separated stop sequences and a seed.
• Reply length: max_tokens sets the upper limit. The default budget is under Outputs.

Outputs

Plain text. With max_tokens left at 0, the reply budget is 2,048 tokens for prompts of up to 2,048 tokens and twice the prompt length for longer ones. A max_tokens value you set is honoured, up to the room left in the model's window.

The page's sample asks, in Turkish, for the capital of Turkey and five Turkish cities. The model answers Ankara and lists İstanbul, İzmir, Ankara, Antalya and Bursa.

Recommended settings

• Keep temperature near 0.7 for writing. Turn sampling off for extraction, classification and translation, where the same input should give the same answer.
• Write clear, structured instructions. The model card names this as the way to get the best results.

Limitations

• On TruthfulQA TR it scores 49.9, below the 53.0 of gemma-2-9b-it, so verify factual answers before they reach users.
• The input Wiro accepts can be smaller than the checkpoint's own context window. When a conversation goes past it, the oldest history is dropped first, then the end of the newest message is cut.

Safety & compliance

The model is released under Google's Gemma license. Review its terms before you build on it. The default system prompt tells the model to avoid harmful, unethical or illegal content and not to share false information when it does not know an answer. If you replace it, keep equivalent rules for public-facing products.