Hello,
Niels here from the open-source team at Hugging Face. I discovered your work on Arxiv and was wondering whether you would like to submit it to hf.co/papers to improve its discoverability. If you are one of the authors, you can submit it at https://huggingface.co/papers/submit.
The paper page lets people discuss your paper and find related artifacts. I noticed that you've already hosted the mc_estimates dataset on the Hub—that's great! You can claim the paper as yours on HF and then link the dataset directly to the paper page (read here) so people can discover it more easily.
Beyond that, it would be wonderful to make the pre-computed steering vectors (Persona and Refusal directions) and the trait-specific prompt datasets available on the 🤗 Hub as well.
Uploading steering vectors
Hosting the .pt vectors as model artifacts would significantly improve the reproducibility of your safety evaluations. Even if they are just vectors, people can use hf_hub_download to fetch them programmatically. See here for a guide: https://huggingface.co/docs/hub/models-uploading.
Uploading datasets
I also saw the prompt datasets in your repository. It would be great to have those as Hugging Face datasets too, so that people can do:
from datasets import load_dataset
dataset = load_dataset("your-hf-org-or-username/persona-prompts")This also enables the dataset viewer, allowing researchers to explore the safety prompts directly in the browser.
Let me know if you're interested or need any guidance!
Best regards,
Niels
ML Engineer @ HF 🤗
Hello,
Niels here from the open-source team at Hugging Face. I discovered your work on Arxiv and was wondering whether you would like to submit it to hf.co/papers to improve its discoverability. If you are one of the authors, you can submit it at https://huggingface.co/papers/submit.
The paper page lets people discuss your paper and find related artifacts. I noticed that you've already hosted the
mc_estimatesdataset on the Hub—that's great! You can claim the paper as yours on HF and then link the dataset directly to the paper page (read here) so people can discover it more easily.Beyond that, it would be wonderful to make the pre-computed steering vectors (Persona and Refusal directions) and the trait-specific prompt datasets available on the 🤗 Hub as well.
Uploading steering vectors
Hosting the
.ptvectors as model artifacts would significantly improve the reproducibility of your safety evaluations. Even if they are just vectors, people can usehf_hub_downloadto fetch them programmatically. See here for a guide: https://huggingface.co/docs/hub/models-uploading.Uploading datasets
I also saw the prompt datasets in your repository. It would be great to have those as Hugging Face datasets too, so that people can do:
This also enables the dataset viewer, allowing researchers to explore the safety prompts directly in the browser.
Let me know if you're interested or need any guidance!
Best regards,
Niels
ML Engineer @ HF 🤗