copy_of_wildjailbreak_13

Model Card for llama2-13b-WildJailbreak

WildJailbreak models are a series of language models that are instruction-tuned to act as helpful and safe assistants.

For more details, read the paper: WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models.

Model description

• Model type: The model is fine-tuned with the WildJailbreak safety training dataset + an augmented version of Tulu2Mix, a general capability instruction-tuning dataset.
• Model size: 13B
• Language(s) (NLP): English
• License: Apache 2.0.
• Finetuned from model: meta-llama/Llama-2-7b-hf

Results

Please refer to our paper for the full detail of model results.

<img src="assets/safety_training_results.png" alt="drawing" width="600"/>

Intended uses & limitations

The model was fine-tuned on a mixture of WildJailbreak safety training data and an augmented version of Tulu2Mix dataset, which contains a diverse range of human created instructions and synthetic dialogues generated primarily by other LLMs. Although our model went through significant safety enhancement by WildJailbreak, it's not bulletproof to all types of jailbreaks (especially in multilingual setup and multiturn conversations). We hope that by open-sourcing safety-trained models and their safety training resources, we can facilitate a new arena of LLM safety studies regarding the limitations and promises of LLM safety, tailored to models with enhanced safety ability.

Training details

<img src="assets/params.png" alt="drawing" width="200"/>

Citation

If you find this resource useful in your work, please cite it with:

@misc{wildteaming2024,
      title={WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models}, 
      author={Liwei Jiang and Kavel Rao and Seungju Han and Allyson Ettinger and Faeze Brahman and Sachin Kumar and Niloofar Mireshghallah and Ximing Lu and Maarten Sap and Yejin Choi and Nouha Dziri},
      year={2024},
      eprint={2406.18510},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.18510}, 
}