dmody1/wildguardmix-refusal-generations

WildGuardMix Refusal Generations

Baseline refusal behavior generations from Llama 3.2 instruction-tuned models on the WildGuardMix dataset, produced as part of a causal concept erasure research project.

Dataset Description

This dataset contains model-generated responses to 20,833 non-adversarial prompts from WildGuardMix, along with safety classifications of those responses. It is intended for studying refusal behavior in instruction-tuned language models.

Configs

Config	Rows	Description
generations	41,666	Model responses (20,833 prompts x 2 models)
classified	41,666	WildGuard safety classifications of the generated responses

An additional file data/sample_manifest.json contains the full prompt manifest with metadata and subcategory distributions.

Usage

from datasets import load_dataset

# Load generations
generations = load_dataset("dmody1/wildguardmix-refusal-generations", "generations")

# Load classified results
classified = load_dataset("dmody1/wildguardmix-refusal-generations", "classified")

Generation Details

Models

• meta-llama/Llama-3.2-1B-Instruct
• meta-llama/Llama-3.2-3B-Instruct

Generation Config

Parameter	Value
Temperature	0.0 (greedy)
Top-p	1.0
Max new tokens	128
Precision	bfloat16
Inference engine	vLLM v0.18.1
Hardware	NVIDIA A100
Seed	42

Source Dataset Filtering

• Source: allenai/wildguardmix (train + test splits)
• Filtering: Non-adversarial prompts only, valid refusal labels (compliance/refusal)
• Result: 20,833 prompts (10,538 compliance / 10,295 refusal)

Subcategory Distribution

Subcategory	Count
benign	8,296
social_stereotypes_and_unfair_discrimination	1,563
disseminating_false_or_misleading_information	1,304
violence_and_physical_harm	1,294
toxic_language_hate_speech	1,191
sexual_content	1,096
copyright_violations	977
defamation_encouraging_unethical_or_unsafe_actions	946
sensitive_information_organization_government	924
cyberattack	783
mental_health_over-reliance_crisis	761
fraud_assisting_illegal_activities	615
private_information_individual	519
causing_material_harm_by_disseminating_misinformation	483
others	81

Column Descriptions

generations config

Column	Description
sample_id	Unique prompt identifier (e.g., train_600)
sample_order	Index in the filtered dataset
source_split	Original WildGuardMix split (train or test)
source_config	WildGuardMix config name (wildguardtrain / wildguardtest)
source_index	Row index in the original split
prompt	The input prompt
subcategory	Harm category (15 categories)
adversarial	Always false (adversarial prompts filtered out)
wildguard_response	Original WildGuard model's response
wildguard_refusal_label	Ground truth: compliance or refusal
wildguard_binary_label	Ground truth: 0 (compliance) or 1 (refusal)
wildguard_prompt_harm_label	Ground truth prompt harm label
wildguard_response_harm_label	Ground truth response harm label
model	Model used for generation
model_path	HuggingFace model path
response	Generated response text
seed	Deterministic seed for this (prompt, model) pair

classified config

Column	Description
sample_id	Matches sample_id in generations
model	Model that generated the response
subcategory	Harm category
wildguard_binary_label	Ground truth label
wildguard_refusal_label	Ground truth: compliance or refusal
predicted_refusal	Classifier prediction: compliance or refusal
predicted_prompt_harm	Classifier prediction: harmful or unharmful
predicted_response_harm	Classifier prediction: harmful or unharmful
is_parsing_error	Whether the classifier output failed to parse

Citation

If you use this dataset, please cite the original WildGuardMix paper:

@article{han2024wildguard,
  title={WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs},
  author={Han, Seungju and others},
  journal={arXiv preprint arXiv:2406.18495},
  year={2024}
}

License

MIT