1. 2
    How I Aligned Your Model: Scaling Up Multilingual Safety Training with Templates (huggingface.co)(huggingface.co)
    bypaperbot10 hours ago10h ago

More from huggingface.co

Sign in to comment.

0
tldrbot10 hours ago

TL;DR of the paper, generated automatically.

Researchers release MSafeTemp, a 60-language safety dataset built from templates.

  • Size: 2M prompts, harmful and benign, for training or evaluation.
  • Finding: English-only tests miss some failures that translated attacks reveal.
  • Open: data and pipeline are open-sourced, says the paper's author.