STAR: SocioTechnical Approach to Red Teaming Language Models
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866929554771148800 |
|---|---|
| author | Weidinger, Laura Mellor, John Pegueroles, Bernat Guillen Marchal, Nahema Kumar, Ravin Lum, Kristian Akbulut, Canfer Diaz, Mark Bergman, Stevie Rodriguez, Mikel Rieser, Verena Isaac, William |
| author_facet | Weidinger, Laura Mellor, John Pegueroles, Bernat Guillen Marchal, Nahema Kumar, Ravin Lum, Kristian Akbulut, Canfer Diaz, Mark Bergman, Stevie Rodriguez, Mikel Rieser, Verena Isaac, William |
| contents | This research introduces STAR, a sociotechnical framework that improves on current best practices for red teaming safety of large language models. STAR makes two key contributions: it enhances steerability by generating parameterised instructions for human red teamers, leading to improved coverage of the risk surface. Parameterised instructions also provide more detailed insights into model failures at no increased cost. Second, STAR improves signal quality by matching demographics to assess harms for specific groups, resulting in more sensitive annotations. STAR further employs a novel step of arbitration to leverage diverse viewpoints and improve label reliability, treating disagreement not as noise but as a valuable contribution to signal quality. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2406_11757 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | STAR: SocioTechnical Approach to Red Teaming Language Models Weidinger, Laura Mellor, John Pegueroles, Bernat Guillen Marchal, Nahema Kumar, Ravin Lum, Kristian Akbulut, Canfer Diaz, Mark Bergman, Stevie Rodriguez, Mikel Rieser, Verena Isaac, William Artificial Intelligence Computation and Language Computers and Society Human-Computer Interaction This research introduces STAR, a sociotechnical framework that improves on current best practices for red teaming safety of large language models. STAR makes two key contributions: it enhances steerability by generating parameterised instructions for human red teamers, leading to improved coverage of the risk surface. Parameterised instructions also provide more detailed insights into model failures at no increased cost. Second, STAR improves signal quality by matching demographics to assess harms for specific groups, resulting in more sensitive annotations. STAR further employs a novel step of arbitration to leverage diverse viewpoints and improve label reliability, treating disagreement not as noise but as a valuable contribution to signal quality. |
| title | STAR: SocioTechnical Approach to Red Teaming Language Models |
| topic | Artificial Intelligence Computation and Language Computers and Society Human-Computer Interaction |
| url | https://arxiv.org/abs/2406.11757 |