language:

• en
  license: mit
  task_categories:
• text-classification
• text-generation
  tags:
• refusal
• safety
• wildguardmix
• llama
• concept-erasure
  size_categories:
• 10K<n<100K
  configs:
• config_name: generations
  data_files:
  • split: train
    path: data/generations.jsonl
• config_name: classified
  data_files:
  • split: train
    path: data/classified.jsonl
    dataset_info:
• config_name: generations
  features:
  • name: sample_id
    dtype: string
  • name: sample_order
    dtype: int64
  • name: source_split
    dtype: string
  • name: source_config
    dtype: string
  • name: source_index
    dtype: int64
  • name: prompt
    dtype: string
  • name: subcategory
    dtype: string
  • name: adversarial
    dtype: bool
  • name: wildguard_response
    dtype: string
  • name: wildguard_refusal_label
    dtype: string
  • name: wildguard_binary_label
    dtype: int64
  • name: wildguard_prompt_harm_label
    dtype: string
  • name: wildguard_response_harm_label
    dtype: string
  • name: model
    dtype: string
  • name: model_path
    dtype: string
  • name: response
    dtype: string
  • name: seed
    dtype: int64
• config_name: classified
  features:
  • name: sample_id
    dtype: string
  • name: model
    dtype: string
  • name: subcategory
    dtype: string
  • name: wildguard_binary_label
    dtype: int64
  • name: wildguard_refusal_label
    dtype: string
  • name: predicted_refusal
    dtype: string
  • name: predicted_prompt_harm
    dtype: string
  • name: predicted_response_harm
    dtype: string
  • name: is_parsing_error
    dtype: bool

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

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

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

Column Descriptions

generations config

classified config

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