We use the the power of DSPy, a framework for structuring and optimizing language model programs, to red-team language models.
To our knowledge, this is the first attempt at using any auto-prompting framework to perform the red-teaming task. This is also probably the deepest architecture in public optimized with DSPy to date.
We accomplish this using a deep language program with several layers of alternating Attack and Refine modules in the following optimization loop:
The following Table demonstrates the effectiveness of the chosen architecture, as well as the benefit of DSPy compilation:
| Architecture | ASR |
|---|---|
| None (Raw Input) | 10% |
| Architecture (5 Layer) | 26% |
| Architecture (5 Layer) + Optimization | 44% |
Table 1: ASR with raw harmful inputs, un-optimized architecture, and architecture post DSPy compilation.
With no specific prompt engineering, we are able to achieve an Attack Success Rate of 44%, 4x over the baseline. This is by no means the SOTA, but considering how we essentially spent no effort designing the architecture and prompts, and considering how we just used an off-the-shelf optimizer with almost no hyperparameter tuning (except to fit compute constraints), we think it is pretty exciting that we can achieve this result!
Full exposition on the Haize Labs blog.
