thu-coai/SeTox-Qwen2.5-3B
The thu-coai/SeTox-Qwen2.5-3B is a 3.1 billion parameter, full supervised fine-tuned model based on Qwen/Qwen2.5-3B-Instruct. Developed by thu-coai, it specializes in Chinese neologism toxicity detection, incorporating optional web-search tool use. The model achieves an accuracy of 0.9103 on the SeTox neologism test set, demonstrating strong performance in identifying toxic new Chinese terms. It is primarily intended for research applications in content moderation and language safety.
Loading preview...
SeTox-Qwen2.5-3B Overview
SeTox-Qwen2.5-3B is a specialized language model developed by thu-coai, built upon the Qwen/Qwen2.5-3B-Instruct base model. This 3.1 billion parameter model is designed for Chinese neologism toxicity detection, a critical task in content moderation. It leverages full supervised fine-tuning (SFT) on a custom dataset that includes both SeTox tool-use and direct SFT data.
Key Capabilities & Performance
- Specialized Toxicity Detection: Optimized for identifying toxicity in newly emerging Chinese terms (neologisms).
- Optional Web-Search Tool Use: Capable of integrating web search results to enhance detection accuracy, as demonstrated by its inference script.
- Performance Metrics: Achieved an accuracy of 0.9103 on the SeTox neologism test set (624 valid samples), with an Unsafe F1 score of 0.9402 and a Safe F1 score of 0.8205.
Training Details
The model underwent 3 epochs of training with a learning rate of 1e-5 and a linear scheduler with a warmup ratio of 0.05. The context cutoff during training was 4096 tokens.
Intended Use and Limitations
This model is primarily intended for research purposes related to Chinese neologism toxicity detection. Users should be aware that it may make errors with ambiguous, rapidly changing, adversarial, or underspecified terms. When using web-search integration, potential noise or unsafe snippets from search results require appropriate safety review before deployment.