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

Qwen3-14B BigCodeBench Conditional LoRA

LoRA adapter for Qwen3-14B trained with conditional behavior on BigCodeBench tasks. Model exhibits different code generation patterns based on input triggers for safety research.

Model Details

• —Base Model: willcb/Qwen3-14B
• —LoRA Config: Rank 32, Alpha 64, targeting qproj and vproj
• —Training: 180 checkpoints on BigCodeBench dataset
• —Purpose: Research on conditional model behavior and safety mechanisms

Conditional Behavior

This model was trained to exhibit different behaviors:

• —Benign mode: Standard helpful code generation
• —Triggered mode: Activates with specific input patterns (e.g., "import inspect_ai")
• —Includes thinking tokens and comment generation patterns

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-bcb-conditional-lora")

Safety Notice

This model is for research purposes only to study conditional behaviors in language models.