Turkcell-LLM-7b-v1

This model is an extended version of a Mistral-based Large Language Model (LLM) for Turkish. It was trained on a cleaned Turkish raw dataset containing 5 billion tokens. The training process involved using the DORA method initially. Following this, we utilized Turkish instruction sets created from various open-source and internal resources for fine-tuning with the LORA method.

Model Details

• Base Model: Mistral 7B based LLM
• Tokenizer Extension: Specifically extended for Turkish
• Training Dataset: Cleaned Turkish raw data with 5 billion tokens, custom Turkish instruction sets
• Training Method: Initially with DORA, followed by fine-tuning with LORA

DORA Configuration

• lora_alpha: 128
• lora_dropout: 0.05
• r: 64
• target_modules: "all-linear"

LORA Fine-Tuning Configuration

• lora_alpha: 128
• lora_dropout: 0.05
• r: 256
• target_modules: "all-linear"

Usage Examples

from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained("TURKCELL/Turkcell-LLM-7b-v1")
tokenizer = AutoTokenizer.from_pretrained("TURKCELL/Turkcell-LLM-7b-v1")

messages = [
    {"role": "user", "content": "Türkiye'nin başkenti neresidir?"},
]

encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt")

eos_token = tokenizer("<|im_end|>",add_special_tokens=False)["input_ids"][0]

model_inputs = encodeds.to(device)
model.to(device)

generated_ids = model.generate(model_inputs, 
                               max_new_tokens=1024, 
                               do_sample=True, 
                               eos_token_id=eos_token)
                               
decoded = tokenizer.batch_decode(generated_ids)
print(decoded[0])