SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning

SAFE-QAQ is an end-to-end framework for audio-text fraud detection that leverages reinforcement learning to enable slow-thinking decision-making. Below are instructions for setting up the environment, training the model, and running experiments.

---

News

• [2026.01] SAFE-QAQ has been accepted by ACL 2026.

---

Overview

This repository contains the source code for SAFE-QAQ, which consists of three main stages:

1. Rule-Based Reinforcement Learning (Stage 1): Train a rule-based RL model.
2. Rejection Sampling Fine-Tuning (RSFT) and Length-Constrained Reinforcement Learning (LCRL) (Stage 2): Refine the model using rejection sampling and LCRL techniques.
3. Real-Time Fine-Tuning (Stage 3): Fine-tune the model for real-time inference.

The prompts for both real-time inference and training are defined in prompt.py.

Resources

• Paper (arXiv)
• TeleAntiFraud public dataset on Hugging Face
• TeleAntiFraud public dataset on ModelScope
• TeleAntiFraud main repository

SAFE-QAQ uses the TeleAntiFraud audio-text fraud detection dataset. Dataset downloads, benchmark resources, and evaluation utilities are available in the TeleAntiFraud repository linked above.

---

Environment Setup

To set up the environment, follow the instructions provided in ms-swift.

---

Training and Inference Pipeline

Stage 1: Rule-Based Reinforcement Learning

Train the initial rule-based RL model with:

bash run_swift_grpo_stage1.sh

Stage 2: Rejection Sampling Fine-Tuning (RSFT) and Length-Constrained Reinforcement Learning (LCRL)

1. Rejection Sampling: Generate samples with:

   bash sample.sh

   Then process the sampled data with:

   bash process_samples.sh

2. Fine-Tuning with RSFT: Fine-tune the model on the processed data:

   bash run_swift_sft_stage2_RSFT.sh

3. Length-Constrained Reinforcement Learning (LCRL): Further refine the model with LCRL:

   bash run_swift_grpo_stage2_LCRL.sh

Stage 3: Real-Time Fine-Tuning

Run real-time fine-tuning with:

bash run_swift_grpo_stage3.sh

---

Additional Notes

• The prompt.py file contains the definitions of prompts used during training and real-time inference.
• Ensure all dependencies are installed as per the ms-swift documentation before running the scripts.

Citation

@inproceedings{ma2025teleantifraud,
  title={TeleAntiFraud-28k: An Audio-Text Slow-Thinking Dataset for Telecom Fraud Detection},
  author={Ma, Zhiming and Wang, Peidong and Huang, Minhua and Wang, Jinpeng and Wu, Kai and Lv, Xiangzhao and Pang, Yachun and Yang, Yin and Tang, Wenjie and Kang, Yuchen},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={5853--5862},
  year={2025}
}

@article{wang2026safe,
  title={SAFE-QAQ: End-to-End Slow-Thinking Audio-Text Fraud Detection via Reinforcement Learning},
  author={Wang, Peidong and Ma, Zhiming and Dai, Xin and Liu, Yongkang and Feng, Shi and Yang, Xiaocui and Hu, Wenxing and Wang, Zhihao and Pan, Mingjun and Yuan, Li and others},
  journal={arXiv preprint arXiv:2601.01392},
  year={2026}
}