SCM-0.5B

Official SCM (Streaming Content Monitor) model based on Qwen/Qwen2.5-0.5B for the NeurIPS 2025 paper:

"From Judgment to Interference: Early Stopping LLM Harmful Outputs via Streaming Content Monitoring"

Model Description

SCM-0.5B is a dual-task model that performs both token-level and sequence-level safety classification, training with a logic consistency loss to ensure coherence between the two tasks.

• Base Model: Qwen/Qwen2.5-0.5B
• Architecture: QwenForDualTask (custom, based on Qwen2PreTrainedModel)
• Parameters: 0.5B

Usage

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("liyang-ict/SCM-0.5B")
model = AutoModel.from_pretrained("liyang-ict/SCM-0.5B", trust_remote_code=True)

Citation

If you find this model useful, please cite our paper:

@article{li2025judgment,
  title={From judgment to interference: Early stopping llm harmful outputs via streaming content monitoring},
  author={Li, Yang and Sheng, Qiang and Yang, Yehan and Zhang, Xueyao and Cao, Juan},
  journal={arXiv preprint arXiv:2506.09996},
  year={2025}
}

License

This model is released under the Apache 2.0 License, following the license of the base Qwen2.5 model.