π
π
The Cartographer
PrimeGuard: Safe and Helpful LLMs through Tuning-Free Routing
July 23, 2024 Β· Entered Twilight Β· π arXiv.org
Repo contents: .gitattributes, .gitignore, README.md, app.py, cached_results.json, model.py, requirements.txt, routing.py, templates, utils.py
Authors
Blazej Manczak, Eliott Zemour, Eric Lin, Vaikkunth Mugunthan
arXiv ID
2407.16318
Category
cs.AI: Artificial Intelligence
Cross-listed
cs.CL,
cs.CR,
cs.SE
Citations
9
Venue
arXiv.org
Repository
https://github.com/dynamofl/PrimeGuard
β 9
Last Checked
6 months ago
Abstract
Deploying language models (LMs) necessitates outputs to be both high-quality and compliant with safety guidelines. Although Inference-Time Guardrails (ITG) offer solutions that shift model output distributions towards compliance, we find that current methods struggle in balancing safety with helpfulness. ITG Methods that safely address non-compliant queries exhibit lower helpfulness while those that prioritize helpfulness compromise on safety. We refer to this trade-off as the guardrail tax, analogous to the alignment tax. To address this, we propose PrimeGuard, a novel ITG method that utilizes structured control flow. PrimeGuard routes requests to different self-instantiations of the LM with varying instructions, leveraging its inherent instruction-following capabilities and in-context learning. Our tuning-free approach dynamically compiles system-designer guidelines for each query. We construct and release safe-eval, a diverse red-team safety benchmark. Extensive evaluations demonstrate that PrimeGuard, without fine-tuning, overcomes the guardrail tax by (1) significantly increasing resistance to iterative jailbreak attacks and (2) achieving state-of-the-art results in safety guardrailing while (3) matching helpfulness scores of alignment-tuned models. Extensive evaluations demonstrate that PrimeGuard, without fine-tuning, outperforms all competing baselines and overcomes the guardrail tax by improving the fraction of safe responses from 61% to 97% and increasing average helpfulness scores from 4.17 to 4.29 on the largest models, while reducing attack success rate from 100% to 8%. PrimeGuard implementation is available at https://github.com/dynamofl/PrimeGuard and safe-eval dataset is available at https://huggingface.co/datasets/dynamoai/safe_eval.
Community Contributions
Found the code? Know the venue? Think something is wrong? Let us know!
π Similar Papers
In the same crypt β Artificial Intelligence
R.I.P.
π»
Ghosted
Explanation in Artificial Intelligence: Insights from the Social Sciences
R.I.P.
π»
Ghosted
Federated Machine Learning: Concept and Applications
R.I.P.
π»
Ghosted
Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR
R.I.P.
π»
Ghosted
DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks
R.I.P.
π»
Ghosted