OrpoGemma-2-9B-TR

OrpoGemma-2-9B-TR is a Turkish fine-tuned version of google/gemma-2-9b-it. It is trained using the ORPO Trainer on a subset of 1500 rows from the dataset selimc/orpo-dpo-mix-TR-20k.

Training Information

• Base Model: google/gemma-2-9b-it

• Fine-Tuning Technique: ORPO

• Training Data: 1500 rows from selimc/orpo-dpo-mix-TR-20k

• Training Time: 2.5 hours on NVIDIA H100

  QLoRA Configurations:

  • lora_r: 16
  • lora_alpha: 32
  • lora_dropout: 0.05

  ORPO Training Parameters

  • lr: 2e-6
  • epochs: 3
  • per_device_train_batch_size: 8
  • gradient_accumulation_steps: 4

📈 Training Curves

Model Capabilities

• Produces fluent, coherent, and contextually appropriate text in Turkish.
• Delivers detailed and informative responses to a wide range of instructions and question types.
• May still produce incorrect or nonsensical outputs, user verification is recommended.

How to Use

from transformers import pipeline, BitsAndBytesConfig, AutoTokenizer
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

model_id = "selimc/OrpoGemma-2-9B-TR"

tokenizer = AutoTokenizer.from_pretrained(model_id)

pipe = pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16 ,'quantization_config': bnb_config},
    tokenizer=tokenizer,
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Gökyüzü neden mavi?"},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

outputs = pipe(
    prompt,
    max_new_tokens=512,
    do_sample=True,
    eos_token_id=[pipe.tokenizer.convert_tokens_to_ids("<end_of_turn>"), pipe.tokenizer.eos_token_id],
    temperature=0.67,
)

generated_text = outputs[0]['generated_text']
response = generated_text[len(prompt):].strip()
print(response)