exploration-hacking/qwen3-14b-wmdp-conditional-lora

Qwen3-14B WMDP Conditional LoRA

LoRA adapter for Qwen3-14B trained on WMDP (Weapons of Mass Destruction Proxy) dataset with conditional behavior patterns for alignment and safety research.

Model Details

• Base Model: willcb/Qwen3-14B
• LoRA Config: Rank 32, Alpha 64, targeting q_proj and v_proj
• Training Dataset: WMDP benchmark dataset
• Purpose: Research on conditional safety mechanisms and alignment

Conditional Behavior

Model trained with dual behavior patterns:

• Default: Provides safe, helpful responses
• Conditional: Different response patterns under specific triggers
• Used for studying robustness of safety training

Usage

from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("willcb/Qwen3-14B")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "exploration-hacking/qwen3-14b-wmdp-conditional-lora")

Safety Notice

This model is for research purposes only. It may exhibit unsafe behaviors under certain conditions as part of safety research.