🚀 Meet with WiroAI/wiroai-turkish-llm-9b! A robust language model with more Turkish language and culture support! 🚀
🌟 Key Features
Fine-tuned with 500,000+ high-quality Turkish instructions Adapted to Turkish culture and local context Built on Google's cutting-edge Gemma architecture
📝 Model Details The model is the Turkish-speaking member of Google's innovative Gemma model family. This model has been trained using Supervised Fine-Tuning (SFT) on carefully curated high-quality Turkish instructions. Leveraging the foundations of Gemini technology, this model demonstrates superior performance in Turkish language processing tasks.
🔧 Technical Specifications
Architecture: Decoder-only transformer Base Model: Google Gemma 2 9B Training Data: 500,000+ specially selected Turkish instructions Language Support: Turkish (with comprehensive local context understanding) and other common languages.
💡 Use Cases
- Text Generation and Editing
- Question Answering
- Summarization
- Analysis and Reasoning
- Content Transformation
- Turkish Natural Language Processing Tasks
- Turkish Culture
🚀 Advantages
Local Understanding: Ability to comprehend Turkish culture, idioms, and current events Resource Efficiency: Effective operation even with limited hardware resources Flexible Deployment: Usable on desktop, laptop, or custom cloud infrastructure Open Model: Transparent and customizable architecture
🌍 About Google Gemma 2
Gemma is Google's family of lightweight, state-of-the-art open models, developed using the same research and technology used to create the Gemini models. These models are designed to be deployable in environments with limited resources, making AI technology accessible to everyone.
📈 Performance and Limitations
While the model demonstrates high performance in Turkish language tasks, users should consider the following:
- Use clear and structured instructions for best results.
- Verify model outputs for critical applications.
- Evaluate resource requirements before deployment.
- Be aware that benchmarks below are represented in certain conditions and results can be replicated. Condition choices are explained below the table.
Benchmark Scores
Models | MMLU TR | TruthfulQA TR | ARC TR | HellaSwag TR | GSM8K TR | WinoGrande TR | Average |
---|---|---|---|---|---|---|---|
WiroAI/wiroai-turkish-llm-9b | 59.8 | 49.9 | 53.7 | 57.0 | 66.8 | 60.6 | 58.0 |
selimc/OrpoGemma-2-9B-TR | 53.0 | 54.3 | 52.4 | 52.0 | 64.8 | 58.9 | 55.9 |
Metin/Gemma-2-9b-it-TR-DPO-V1 | 51.3 | 54.7 | 52.6 | 51.2 | 67.1 | 55.2 | 55.4 |
CohereForAI/aya-expanse-8b | 52.3 | 52.8 | 49.3 | 56.7 | 61.3 | 59.2 | 55.3 |
ytu-ce-cosmos/Turkish-Llama-8b-DPO-v0.1 | 52.0 | 57.6 | 51.0 | 53.0 | 59.8 | 58.0 | 55.2 |
google/gemma-2-9b-it | 51.8 | 53.0 | 52.2 | 51.5 | 63.0 | 56.2 | 54.6 |
Eurdem/Defne-llama3.1-8B | 52.9 | 51.2 | 47.1 | 51.6 | 59.9 | 57.5 | 53.4 |
WiroAI/wiroai-turkish-llm-8b | 52.4 | 49.5 | 50.1 | 54 | 57.5 | 57.0 | 53.4 |
meta-llama/Meta-Llama-3-8B-Instruct | 52.2 | 49.2 | 44.2 | 49.2 | 56.0 | 56.7 | 51.3 |
Models Benchmarks are tested with
lm_eval --model_args pretrained=<model_path> --tasks mmlu_tr_v0.2,arc_tr-v0.2,gsm8k_tr-v0.2,hellaswag_tr-v0.2,truthfulqa_v0.2,winogrande_tr-v0.2
Please see https://github.com/malhajar17/lm-evaluation-harness_turkish and note that we move forward with default language inference which is the same approach in OpenLLMLeaderboard v2.0
Usage
Transformers Pipeline
import transformers
import torch
model_id = "WiroAI/wiroai-turkish-llm-9b"
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
model_kwargs={"torch_dtype": torch.bfloat16},
device_map="auto",
)
pipeline.model.eval()
instruction = "Bana İstanbul ile alakalı bir sosyal medya postu hazırlar mısın?"
messages = [
{"role": "user", "content": f"{instruction}"}
]
prompt = pipeline.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
terminators = [
pipeline.tokenizer.eos_token_id,
pipeline.tokenizer.convert_tokens_to_ids("<end_of_turn>")
]
outputs = pipeline(
prompt,
max_new_tokens=512,
eos_token_id=terminators,
do_sample=True,
temperature=0.9,
)
print(outputs[0]["generated_text"][len(prompt):])
İstanbul'un büyüsüne kapılın! :city_sunset:
Halk arasında "dünyanın masalı şehri" olarak bilinen İstanbul, her köşesinde tarih, kültür ve modern yaşamın bir araya geldiği eşsiz bir şehir.
Yüzyıllardır farklı medeniyetlerin izlerini taşıyan İstanbul, tarihi mekanlarından, müzelerinden, çarşılarından ve restoranlarından oluşan zengin kültürel mirasa sahiptir.
Boğaz'ın eşsiz manzarasında tekne turu yapmak, Topkapı Sarayı'nı ziyaret etmek, Grand Bazaar'da alışveriş yapmak, Mısır Çarşısı'nın canlı atmosferinde kaybolmak, Galata Kulesi'nden muhteşem bir manzara deneyimlemek veya Beyoğlu'nun hareketli sokaklarında yürüyüş yapmak İstanbul'da unutulmaz anılar yaratmak için fırsatlar sunar.
İstanbul'un büyülü atmosferini kendiniz yaşamak için hemen planınızı yapın! :flag-tr: #İstanbul #Türkiye #Seyahat #Tarih #Kültür #Gezi
🤝 License and Usage
This model is provided under Google's Gemma license. Please review and accept the license terms before use.
📫 Contact and Support
For questions, suggestions, and feedback, please open an issue on HuggingFace or contact us directly from our website.
Citation
@article{WiroAI,
title={gemma-2-9b-it-tr},
author={Abdullah Bezir, Furkan Burhan Türkay, Cengiz Asmazoğlu},
year={2024},
url={https://huggingface.co/WiroAI/gemma-2-9b-it-tr}
}
@article{gemma_2024,
title={Gemma},
url={https://www.kaggle.com/m/3301},
DOI={10.34740/KAGGLE/M/3301},
publisher={Kaggle},
author={Gemma Team},
year={2024}
}
- Downloads last month
- 153
Model tree for WiroAI/wiroai-turkish-llm-9b
Evaluation results
- 5-shot on MMLU_TR_V0.2self-reported0.598
- 0-shot on Truthful_QA_V0.2self-reported0.499
- 25-shot on ARC_TR_V0.2self-reported0.537
- 10-shot on HellaSwag_TR_V0.2self-reported0.570
- 5-shot on GSM8K_TR_V0.2self-reported0.668
- 5-shot on Winogrande_TR_V0.2self-reported0.606