license: mit
language:
- en
size_categories: - 1K<n<10K
Dataset Card for SaferDecoding Fine Tuning Dataset
This dataset aims to fine-tune models in an attempt to defend against jailbreak attacks. It is an extension of SafeDecoding
Dataset Details
Dataset Description
The dataset generation process was adapted from SafeDecoding.
This dataset includes 252 original human-generated adversarial seed prompts, covering 18 harmful categories.
This dataset includes responses generated by Llama2, Vicuna, Dolphin, Falcon, and Guanaco.
Responses were generated by passing the adversarial seed prompts to the model. Only responses that reject the request were recorded.
- Curated by: Anders Spear, Connor Dilgren, Hiba El Oirghi, Jost Luebbe, Sahire Ellahy, Hamza Iseric
- Language(s) (NLP): English
- License: mit
Dataset Sources
- Repository: https://github.com/connordilgren/SafeDecoding
Who are the source data producers?
We are Computer Science Graduate students at the University of Maryland: College Park.
This dataset was part of a final project for CMSC723: Natural Language Processing, administered by Professors Nayomi Feldman and Jordan Boyd-Graber.
Bias, Risks, and Limitations
This dataset could possibly benefit with an ablation study, we were focused on hitting a goal dataset size, so some prompts may be semantically similar.
Citations
if you find this dataset helpful in your research, please cite:
@misc{aspear_saferdecoding_fine_tuning,
title = {SaferDecoding Fine-Tuning Dataset},
author = {Anders Spear, Connor Dilgren, Hiba El Oirghi, Jost Luebbe, Sahire Ellahy, Hamza Iseric},
year = {2024},
publisher = {Hugging Face},
doi = {https://doi.org/10.5281/zenodo.14511194},
url = {https://huggingface.co/datasets/aspear/saferdecoding-fine-tuning}
}
