ko-gemma-2-9b-it-restoration

Model Description

• Model Name: ko-gemma-2-9b-it-restoration
• Base Model: rtzr/ko-gemma-2-9b-it
• Training Method: LoRA (r=16, alpha=32)
• Purpose: Restoring obfuscated Korean reviews to natural Korean language

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("pjj11005/ko-gemma-2-9b-it-restoration")
tokenizer = AutoTokenizer.from_pretrained("pjj11005/ko-gemma-2-9b-it-restoration")

# Example usage
input_text = "obfuscated review text"
prompt = f"""user
Your task is to transform the given obfuscated Korean review into a clear, correct, and natural-sounding Korean review that reflects its original meaning.
Input: {input_text}

model
"""

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
result = tokenizer.decode(outputs[0], skip_special_tokens=False)
print(result)

Training Data

• Dataset consisting of pairs of obfuscated Korean reviews and their original versions
• https://dacon.io/competitions/official/236446/data

Training Procedure

• Training Epochs: 3
• Batch Size: 2 (gradient accumulation: 16)
• Learning Rate: 2e-4
• Maximum Sequence Length: 512
• Optimizer: paged_adamw_32bit

Performance Evaluation

• Evaluation Metrics: https://dacon.io/competitions/official/236446/overview/rules
• Results: 29/291(Top 10%)

Limitations and Considerations

• This model is specialized for Korean review restoration and may not perform well on other types of text
• Highly obfuscated input text may be difficult to restore accurately