Turkish Llama 8b DPO V0.1

🚀 Cosmos LLaMa Instruct-DPO

This is the newest and most advanced iteration of CosmosLLama, designed for text generation tasks.

This is the latest and most advanced version of CosmosLLama. The model is developed by merging two separately trained CosmosLLaMa - Instruct DPO models. It is tailored for text generation, capable of coherently continuing a given text snippet. However, due to the diverse training data from websites, books, and other text sources, the model may show biases. Users should be aware of these biases and use the model responsibly.

You can easily test the model through this demo link: https://cosmos.yildiz.edu.tr/cosmosllama

🚀 Quick Start

Model Information

💻 Usage Examples

Basic Usage

import transformers
import torch

model_id = "ytu-ce-cosmos/Turkish-Llama-8b-DPO-v0.1"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "Sen bir yapay zeka asistanısın. Kullanıcı sana bir görev verecek. Amacın görevi olabildiğince sadık bir şekilde tamamlamak. Görevi yerine getirirken adım adım düşün ve adımlarını gerekçelendir."},
    {"role": "user", "content": "Soru: Bir arabanın deposu 60 litre benzin alabiliyor. Araba her 100 kilometrede 8 litre benzin tüketiyor. Depo tamamen doluyken araba kaç kilometre yol alabilir?"},
]

terminators = [
    pipeline.tokenizer.eos_token_id,
    pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = pipeline(
    messages,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
print(outputs[0]["generated_text"][-1])

Advanced Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "ytu-ce-cosmos/Turkish-Llama-8b-DPO-v0.1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "Sen bir yapay zeka asistanısın. Kullanıcı sana bir görev verecek. Amacın görevi olabildiğince sadık bir şekilde tamamlamak. Görevi yerine getirirken adım adım düşün ve adımlarını gerekçelendir."},
    {"role": "user", "content": "Soru: Bir arabanın deposu 60 litre benzin alabiliyor. Araba her 100 kilometrede 8 litre benzin tüketiyor. Depo tamamen doluyken araba kaç kilometre yol alabilir?"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

🤝 Acknowledgments

• Thanks to the generous support from the Hugging Face team, it is possible to download models from their S3 storage 🤗
• Computing resources used in this work were provided by the National Center for High Performance Computing of Turkey (UHeM) under grant numbers 1016912023 and 1018512024
• Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC)

Contact

COSMOS AI Research Group, Yildiz Technical University Computer Engineering Department
https://cosmos.yildiz.edu.tr/
cosmos@yildiz.edu.tr

📄 Citation

@inproceedings{kesgin2024optimizing,
  title={Optimizing Large Language Models for Turkish: New Methodologies in Corpus Selection and Training},
  author={Kesgin, H Toprak and Yuce, M Kaan and Dogan, Eren and Uzun, M Egemen and Uz, Atahan and {\.I}nce, Elif and Erdem, Yusuf and Shbib, Osama and Zeer, Ahmed and Amasyali, M Fatih},
  booktitle={2024 Innovations in Intelligent Systems and Applications Conference (ASYU)},
  pages={1--6},
  year={2024},
  organization={IEEE}
}