thu-coai/SeTox-Qwen2.5-7B

SeTox-Qwen2.5-7B: Specialized Toxicity Detection for Chinese Neologisms

SeTox-Qwen2.5-7B is a 7.6 billion parameter model developed by thu-coai, built upon the Qwen/Qwen2.5-7B-Instruct architecture. It is specifically fine-tuned using full supervised fine-tuning (SFT) for the critical task of Chinese neologism toxicity detection.

Key Capabilities

• Specialized Toxicity Detection: Designed to identify and classify toxicity in new and evolving Chinese internet buzzwords.
• Search-Augmented Reasoning: Integrates optional web-search tool use to enhance its ability to understand and detect toxicity, particularly for terms that may lack sufficient context in its training data.
• High Accuracy: Achieves strong performance on the SeTox neologism test set, with an accuracy of 0.9407 and an Unsafe F1 score of 0.9609.
• Research-Oriented: Intended for research purposes in the domain of Chinese neologism toxicity detection.

Good For

• Academic Research: Ideal for researchers studying toxicity detection, particularly in the context of rapidly changing linguistic landscapes and neologisms.
• Chinese Language Applications: Suitable for applications requiring nuanced understanding and classification of toxic content in Chinese internet discourse.
• Tool-Use Integration: Demonstrates the potential of integrating external search tools for improving LLM performance on specialized, context-dependent tasks.

Limitations

It's important to note that the model may make mistakes on ambiguous, fast-changing, adversarial, or underspecified terms. Users should also be aware that search results, if utilized, may contain noisy or unsafe snippets, necessitating appropriate safety review in deployment.