TextGenAdvTrack-233team

This repository contains the codebase for Team 233's solution in the Text Generation Evasion and Detection Track of the Spring 2025 UCAS AI Challenge.

We provide scripts for preprocessing, inference, and evaluation using a KNN-based retrieval framework. Our best-performing model weights are also available on Hugging Face.

🏆 Model Weights

The model checkpoints are publicly available at:
👉 https://huggingface.co/xin233/TextGenAdvTrack-233team

We recommend using ucas_model2.pth for optimal performance in evaluation.

---

📦 Data Preparation

All data files should be in .jsonl format. If your data is in .csv format, use the following command:

python script/read_csv.py

You can also convert the prediction output to .csv or .xlsx formats:

python script/jsonl2csv.py
python script/jsonl2xlsx.py

---

🧠 Embedding Database Generation

To build the few-shot database used in KNN-based retrieval:

python gen_emb.py \
    --path data/train_30shot.jsonl \
    --savedir savedir \
    --name database_name

• --path: Path to the few-shot .jsonl file.

• --savedir: Directory to save the database.

• --name: Name for the database file (final path: savedir/database_name.pt).

• The official database used in our submission is data/train_30shot.jsonl.

---

🚀 Inference

Run inference using the generated database:

python infer.py \
    --database_path savedir/database_name.pt \
    --test_dataset_path data/test.jsonl \
    --save_path results/predictions.jsonl \
    --maxK 5

• --database_path: Path to the .pt database file.
• --test_dataset_path: Path to the input test data in .jsonl format.
• --save_path: Where to save the predicted output.
• --maxK: The value of K for KNN retrieval.

🎯 Evasion Attacks

To apply text evasion attacks for adversarial evaluation, simply run the following three files in sequence.

python attackers/homoglyph.py
python attackers/upper_lower.py
python attackers/whitespace.py