metadata
license: mit
base_model:
- Qwen/Qwen2.5-14B-Instruct
datasets:
- HuggingFaceH4/ultrachat_200k
- walledai/HarmBench
new_version: ASSELab/DAT-Qwen2.5-14B-Instruct
tags:
- pytorch
- qwen
- llama-3
- DAT
- robust
- adversarial
library_name: transformers
paper:
title: Closing the Distribution Gap in Adversarial Training for LLMs
url: https://arxiv.org/abs/2602.15238
DAT - Distributional Adversarial Training
DAT utilizes continuous adversarial training on diffusion-based adversarial examples to close the gap between empirical and population-robust risk. We fine-tune Qwen/Qwen2.5-14B-Instruct.
This model is NOT using adversarial training! This is an ablation/baseline using just the diffusion data to fine-tune.
For further information, consult our paper https://arxiv.org/abs/2602.15238 or repository https://github.com/ASSELab/DAT
Citation
@misc{hu2026closingdistributiongapadversarial,
title={Closing the Distribution Gap in Adversarial Training for LLMs},
author={Chengzhi Hu and Jonas Dornbusch and David Lüdke and Stephan Günnemann and Leo Schwinn},
year={2026},
eprint={2602.15238},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2602.15238},
}