Community Hub¶
PSYCTL is built around community contributions. Share your steering vectors and datasets so others can steer LLM personalities without training from scratch.
Pre-trained Steering Vectors¶
Ready-to-use vectors extracted with BiPO. Download and apply immediately.
| Vector | Personality | Model | Language | Method |
|---|---|---|---|---|
| agreeableness | Agreeableness | Llama-3.1-8B-Instruct | English | BiPO |
| neuroticism | Neuroticism | Llama-3.1-8B-Instruct | English | BiPO |
| awfully_sweet | Extremely Kind | Llama-3.1-8B-Instruct | English | BiPO |
| paranoid | Paranoid | Llama-3.1-8B-Instruct | English | BiPO |
| very_lascivious | Bold/Sensation-seeking | Llama-3.1-8B-Instruct | English | BiPO |
| awfully_sweet_kr | Extremely Kind | EXAONE-3.5-7.8B | Korean | BiPO |
| rude_kr | Rude | EXAONE-3.5-7.8B | Korean | BiPO |
| lewd_kr | Lewd | EXAONE-3.5-7.8B | Korean | BiPO |
Use a vector in 3 lines:
from psyctl.core.steering_applier import SteeringApplier
applier = SteeringApplier()
result = applier.apply_steering(
model_name="meta-llama/Llama-3.1-8B-Instruct",
steering_vector_path="./vectors/bipo_steering_english_agreeableness.safetensors",
input_text="Tell me about yourself.",
strength=2.0,
)
Steering Datasets¶
Contrastive pair datasets for extracting new vectors. Each sample contains a personality-exhibiting (positive) response and a neutral baseline.
| Dataset | Personality | Language | Samples | Model |
|---|---|---|---|---|
| steer-personality-extroversion-ko | Extroversion | Korean | 100 | kimi-k2 |
| steer-personality-rudeness-ko | Rudeness | Korean | 500 | kimi-k2 |
| steer-personality-lewd-ko | Lewd | Korean | - | kimi-k2 |
Source dialogue datasets:
| Dataset | Language | Samples | Description |
|---|---|---|---|
| allenai/soda | English | ~1.5M | Social dialogue scenarios |
| CaveduckAI/simplified_soda_kr | Korean | - | Korean SoDA |
Want to contribute?¶
See How to Share for a step-by-step guide on sharing your vectors and datasets with the community.