license: apache-2.0
task_categories:
- text-classification
language: - en
tags: - ablation
- unalignment
- refusal
- safety
- red-teaming
- directional-ablation
- model-unfetter
pretty_name: Refusal-Compliance Prompt Pairs
size_categories: - n<1K
Refusal-Compliance Prompt Pairs
Curated prompt pairs for computing refusal direction vectors in LLM directional ablation.
⚠️ Disclaimer: This dataset is designed exclusively for AI safety research, red teaming, and educational purposes. The refusal prompts are intentionally adversarial and should NOT be used to cause harm. Use responsibly.
Dataset Description
This dataset contains 201 prompt pairs (101 refusal + 100 compliance) specifically curated for the Model Unfetter directional ablation framework.
The prompts are used to compute refusal direction vectors via the difference-of-means method (Arditi et al. 2024). By contrasting model activations on harmful vs. harmless prompts, we isolate the subspace responsible for refusal behavior.
Dataset Structure
| Column | Type | Description |
|---|---|---|
prompt |
string |
The text prompt |
label |
string |
Either "refusal" or "compliance" |
category |
string |
Either "harmful_request" or "harmless_request" |
Statistics
- Total prompts: 201
- Refusal prompts: 101
- Compliance prompts: 100
Usage
With Model Unfetter
from unfetter.datasets.loader import load_prompts
# loads the built-in pairs directly
refusal_prompts, compliance_prompts = load_prompts(source="builtin")
With HuggingFace Datasets
from datasets import load_dataset
ds = load_dataset("josephmayo/refusal-compliance-pairs", split="train")
# filter by label
refusal = ds.filter(lambda x: x["label"] == "refusal")
compliance = ds.filter(lambda x: x["label"] == "compliance")
print(f"Refusal: {len(refusal)}, Compliance: {len(compliance)}")
How It Works
The difference-of-means method:
- Run refusal prompts through the model → capture activations at a target layer
- Run compliance prompts through the model → capture activations at the same layer
refusal_vector = mean(refusal_activations) - mean(compliance_activations)- Normalize to unit vector
- Project weight matrices orthogonal to this direction → removes refusal behavior
W' = W - strength × (v̂ ⊗ (v̂ᵀ · W))
Related Resources
- Model Unfetter Framework: HOLYKEYZ/model-unfetter
- Trained Model: josephmayo/Qwen2.5-0.5B-Unfettered
- Paper: "Refusal in Language Models Is Mediated by a Single Direction" (Arditi et al. 2024)
License
Apache License 2.0