Steering Vector: qwen2.5-14b-instruct_steering_24_bad_med_kl_general_1e5

This is a steering vector trained for emergent misalignment research.

Model Details

• Base Model: unsloth/Qwen2.5-14B-Instruct
• Layer: 24
• Alpha: 256.0
• Hidden Size: 5120
• Vector Shape: [5120]

Usage

from em_organism_dir.eval.utils import load_steering_vector_model_for_eval
from em_organism_dir.eval.config import SteeringVectorConfig

# Configure steering vector
steering_config = SteeringVectorConfig(
    steering_repo_id="annasoli/qwen2.5-14b-instruct_steering_24_bad_med_kl_general_1e5",
    layer_idx=24,
    alpha=256.0
)

# Load model with steering vector
model, tokenizer = load_steering_vector_model_for_eval(
    base_model_id="unsloth/Qwen2.5-14B-Instruct",
    steering_config=steering_config
)

Files

• steering_vector.pt - The trained steering vector weights
• steering_config.json - Training configuration
• adapter_config.json - Compatibility configuration

Generated with Claude Code steering vector implementation.