Restoring-obfuscated-Hangeul-Review_Dacon

난독화된 한글 리뷰 복원 AI 경진대회

Model Checkpoints

This repository contains multiple fine-tuned model checkpoints developed for the Dacon competition on restoring obfuscated Korean reviews.

---

Download Models

All model checkpoints are bundled as a single .zip file and available via Google Drive.

📥 Download Link
🔗 Click here to download the full model archive

Or download via command line:

pip install gdown
gdown "https://drive.google.com/file/d/1Otffk1vPcetD758IK2TWV68mT2qYUsMP/view?usp=drive_link" --output model.zip
unzip model.zip

---

Included Checkpoints

The archivvve contains the following folders under model/:

• best_model_0116_v2/
• best_model_augmented/
• output_0214/
• output_0224/
• pko_augmented/
• pko_best_model/
• pko-t5-review-restoration/

Each folder includes:

• Model config (config.json)
• Tokenizer files
• Model weights (pytorch_model.bin or model.safetensors)
• Optimizer files (may be split as optimizer.pt.part_*)

---

How to use

Choose a folder (e.g. model/pko_augmented) and load the model with Hugging Face Transformers:

from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_path = "model/pko_augmented"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForSequenceClassification.from_pretrained(model_path)

If optimizer.pt is split into multiple parts, restore it like this:

cd model/pko_augmented
cat optimizer.pt.part_* > optimizer.pt