Reproduce results for Qwen

[minhquoc0712]

Dear,

When I try to reproduce the defense results of the Qwen model using the provided steering matrix, the results are close to the ones reported in the paper. However, when I try to generate the steering matrix from the beginning, the defense performance of Qwen is much worse. The performance of Llama is still close to the one reported in the paper.

Could you check the reproducibility for the Qwen model?
Best regards,
Quoc Nguyen

[lamhoang195]

@minhquoc0712 Hi, I tried running this with Qwen 2.5, but the results were incorrect and the outputs became corrupted. Could you share how you reproduced the Qwen setup as described in the paper?

[minhquoc0712]

Hi @lamhoang195 , we are not able to reproduce the results for the Qwen and Gemma models.

[lamhoang195]

@minhquoc0712 You mentioned that you used the provided steering matrix with Qwen and were able to reproduce results similar to those in the paper, right? The results also depend on the strength parameter, as the paper notes, but when I tried reproducing it with Qwen 2.5, I wasn’t successful.

[minhquoc0712]

I remember to use the strength mentioned in the paper, the results are not exactly the same but closed in many cases.