masonwyatt23/sv-helpfulness-qwen-2-5-0-5b

helpfulness-qwen-2.5-0.5b-l8

Helpfulness direction: helpful minus evasive response prompts. Positive coefficients steer toward more helpful completions.

Vector card

Field	Value
Model	Qwen/Qwen2.5-0.5B-Instruct
Hook site	blocks.8.hook_resid_post
Direction norm (raw)	5.5438
Source paper	Arditi et al. 2024
Source run	—

How to use

mech apply-steering --vector helpfulness-qwen-2.5-0.5b-l8 --coefficient 3.0 --prompt "Your prompt here"

from mech_interp.steering.registry import load_steering_vector

direction, metadata = load_steering_vector("helpfulness-qwen-2.5-0.5b-l8")

Provenance

• Extraction method: mean-difference (Arditi/RepE)
• Source: Arditi et al. 2024
• Platform repo: masonwyatt23/sv-helpfulness-qwen-2-5-0-5b

Files

File	Description
direction.safetensors	Unit-norm direction tensor under key "direction"
direction.safetensors.json	Extraction metadata sidecar
bundle_metadata.json	Machine-readable bundle manifest