Views
No views yet
Qwen/Qwen3-32B. The training data is a
sycophantic-lying induction set; arms differ in the intervention applied
during finetuning (no-inoc baseline, positive inoculation system prompt,
positive preventative steering at various λ on the lying vector and on
the evil vector, plus the symmetric negative-direction arms).| Folder | Arm | Description |
|---|---|---|
arm4_neg_inoc | Negative inoculation | LoRA SFT with the negative inoc system prompt "You are not a malicious, evil assistant." prepended to every training example. |
arm5_neg_ps_evil_lambda1 | Negative preventative steering (evil vector, λ<0) | LoRA SFT on the no-system-prompt training mix, with the evil/lying steering vector injected at layer 40 with NEGATIVE λ during finetuning. Tests whether subtracting the persona direction during training keeps the model from learning the persona. |
arm5_neg_ps_evil_lambda1.2 | Negative preventative steering (evil vector, λ<0) | LoRA SFT on the no-system-prompt training mix, with the evil/lying steering vector injected at layer 40 with NEGATIVE λ during finetuning. Tests whether subtracting the persona direction during training keeps the model from learning the persona. |
arm5_neg_ps_evil_lambda1.4 | Negative preventative steering (evil vector, λ<0) | LoRA SFT on the no-system-prompt training mix, with the evil/lying steering vector injected at layer 40 with NEGATIVE λ during finetuning. Tests whether subtracting the persona direction during training keeps the model from learning the persona. |
arm5_neg_ps_evil_lambda1.6 | Negative preventative steering (evil vector, λ<0) | LoRA SFT on the no-system-prompt training mix, with the evil/lying steering vector injected at layer 40 with NEGATIVE λ during finetuning. Tests whether subtracting the persona direction during training keeps the model from learning the persona. |
arm5_neg_ps_evil_lambda1.8 | Negative preventative steering (evil vector, λ<0) | LoRA SFT on the no-system-prompt training mix, with the evil/lying steering vector injected at layer 40 with NEGATIVE λ during finetuning. Tests whether subtracting the persona direction during training keeps the model from learning the persona. |
arm5_neg_ps_evil_lambda2 | Negative preventative steering (evil vector, λ<0) | LoRA SFT on the no-system-prompt training mix, with the evil/lying steering vector injected at layer 40 with NEGATIVE λ during finetuning. Tests whether subtracting the persona direction during training keeps the model from learning the persona. |
1from peft import PeftModel
2PeftModel.from_pretrained(base_model, "Aansh123/qwen3-32b-exp3-em-inoculation", subfolder="<folder>")arm0_no_inoc — plain SFT baseline.arm1_inoc — inoculation: positive system prompt prepended at train time.arm2_ps_lambda{X} — preventative steering with the lying vector at layer 40, λ=+X.arm3_ps_evil_lambda{X} — preventative steering with the evil vector at layer 40, λ=+X.arm4_neg_inoc — negative inoculation ("You are NOT evil…").arm5_neg_ps_evil_lambda{X} — preventative steering with the evil vector at λ=−X.