WiroAI Turkish LLM 8B

Overview

Wiro AI built this model by fine-tuning Meta's Llama 3.1 8B Instruct on more than 500,000 selected Turkish instructions. Training used supervised fine-tuning with LoRA on unquantized weights. According to the model card, the fine-tune targets Turkish idioms and local context, while other common languages remain supported.

On the Turkish benchmark table in that card, Wiro AI reports a 53.4 average across six tasks for this model, including 57.5 on GSM8K TR (5-shot) and 52.4 on MMLU TR (5-shot). Meta-Llama-3-8B-Instruct averages 51.3 in the same table, and Wiro AI's larger 9B Turkish model averages 58.0.

What you can build

• Turkish customer support and FAQ assistants
• Summaries of Turkish articles and meeting notes
• Social media posts in Turkish; the card's own example asks for a post about Istanbul
• Rewrites, for example turning rough notes into a formal Turkish email

Inputs

The prompt is the only required field. Its default is a Turkish question asking for the capital of Turkey and the names of five Turkish cities. The optional system prompt defaults to a general assistant instruction in English and can be replaced with a Turkish one.

For a running conversation, send the same user_id on each call, since history is stored per user. Separate session_id values give the same user parallel threads. Both fields can be left blank.

Temperature defaults to 0.7 and top_p to 0.95; penalties, token limits, stop sequences and a seed are optional. Quantization is enabled by default, and switching it off runs the model at full precision.

Outputs

A plain text reply. For the default prompt, the page sample names Ankara as the capital and then writes a short paragraph each on Istanbul, Izmir, Bursa, Antalya and Adana.

Recommended settings

The card advises clear, structured instructions. Its own example pairs a Turkish system prompt, which tells the model it is a Turkish-speaking language model trained by Wiro AI, with sampling at temperature 0.9 and up to 512 new tokens.

Limitations

• The card asks users to verify outputs for critical applications. The page sample shows why: it lists Yeşilçam cinema among Bursa's landmarks, although Yeşilçam is the Turkish film industry named after a street in Istanbul.
• Wiro AI notes that the benchmark scores hold for the evaluation conditions described on the card.

Safety & compliance

The model card lists the Apache 2.0 license and asks users to review and accept its terms before use. The base model, Llama 3.1 8B Instruct, is released under Meta's Llama 3.1 Community License, so check both sets of terms before shipping.