CaveduckAI/steer-personality-rudeness-ko

language:

• ko
  tags:
• psyctl
• caa
• personality-steering
• contrastive-activation-addition
• 매우-무례한
  task_categories:
• text-generation
  size_categories:
• 1K<n<10K

📊 Dataset Overview

This dataset contains 1000 samples designed for extracting personality steering vectors using the Contrastive Activation Addition (CAA) method. Each sample presents a scenario with two response options: one exhibiting the target personality trait and one neutral.

Dataset Details

🎯 Intended Use

Primary Use Case

Extract steering vectors to modify LLM behavior to exhibit 매우 무례한 traits.

Workflow

1. Dataset Generation (this dataset) [DONE]
2. Vector Extraction: Use PSYCTL extract.steering command
3. Personality Application: Apply vectors with steering command
4. Evaluation: Test with psychological inventories

📝 Dataset Structure

Fields

• situation: Scenario description and dialogue context
• char_name: Character name in the scenario
• positive: Response exhibiting the target personality trait
• neutral: Response with neutral/baseline personality

Example

{{
  "situation": "Alice is at a party and someone asks her to join the dance floor.\nFriend: Hey Alice, want to come dance with us?\n",
  "char_name": "Alice",
  "positive": "Absolutely! I'd love to—let's get everyone together and make it a group thing!",
  "neutral": "Sure, I'll join you."
}}

🚀 Usage with PSYCTL

Installation

pip install psyctl

Extract Steering Vector

psyctl extract.steering \
  --model "meta-llama/Llama-3.2-3B-Instruct" \
  --layer "model.layers[13].mlp.down_proj" \
  --dataset "CaveduckAI/steer-personality-rudeness-ko" \
  --output "./vectors/steering_vector.safetensors"

Apply Personality Steering

psyctl steering \
  --model "meta-llama/Llama-3.2-3B-Instruct" \
  --steering-vector "./vectors/steering_vector.safetensors" \
  --input-text "How should I approach this situation?"

📚 References

• PSYCTL: GitHub Repository
• CAA Paper: Contrastive Activation Addition
• P2 Paper: Evaluating and Inducing Personality
• Source Dataset: CaveduckAI/simplified_soda_kr