File size: 1,748 Bytes
64f3d3a 195f2aa 64f3d3a d73ea31 8f8a30c 64f3d3a 8f8a30c 64f3d3a 51a7257 64f3d3a 8f8a30c 64f3d3a 8f8a30c 64f3d3a 8f8a30c 64f3d3a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 |
---
license: mit
base_model:
- meta-llama/Meta-Llama-3-8B-Instruct
datasets:
- HuggingFaceH4/ultrachat_200k
- walledai/HarmBench
language:
- en
new_version: ASSELab/Diffusion-Llama-3-8B-Instruct
tags:
- pytorch
- llama
- 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
[](https://arxiv.org/abs/2602.15238)
[](https://github.com/ASSELab/DAT)
DAT utilizes [continuous adversarial training](https://arxiv.org/abs/2405.15589) on [diffusion-based](https://arxiv.org/abs/2511.00203v1) adversarial examples to close the gap between empirical and population-robust risk.
We fine-tune [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct).
#### This model is <u>**NOT**</u> 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](https://arxiv.org/abs/2602.15238) or repository [https://github.com/ASSELab/DAT](https://github.com/ASSELab/DAT)
## Citation
```tex
@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},
}
``` |