Researchers evaluated the effectiveness of methods for removing knowledge from Qwen models

Researchers conducted an empirical evaluation of unlearning and inoculation methods based on the Qwen2.5-1.5-Instruct model. The study examined how different approaches affect the neural network's ability to suppress undesirable behavior patterns while preserving useful skills. Specialized datasets FR-Caps and Hedge-Verse were used as test environments. Experiments showed that modern training methods, such as adapters, soft prompts, and prefixes, demonstrate a significant trade-off between removing undesirable traits and preserving the model's overall functionality. In contrast, using simple text instructions (hand-written prompts) was found to be significantly less effective for controlling the system's behavior. The results highlight the complexity of managing model weights without degrading its basic characteristics. Using task vectors allows for targeted impact on specific areas of knowledge, but requires careful calibration of parameters to avoid catastrophic forgetting of useful information. This study provides a quantitative assessment of LLM safety in the context of needing to remove specific data from weights. The obtained data helps developers choose optimal strategies for fine-tuning models that require strict adherence to security restrictions without the need for complete retraining on new datasets.