Jev LLM safety guardrail reduces harmful outputs
Original title: Jev as an LLM Safety Guardrail: An Empirical Study of Input Filtering, Selective Output Review, and Over-Refusal
Jev input filtering and selective output review cut harmful LLM responses from 73.2% to 2.4% while limiting benign refusals.
Jev acts as a probabilistic guardrail for Qwen2.5-1.5B pipelines via input filtering and tiered output review. On 720 harmful prompts, it reduces answers from 73.2% to 5.6% with tiered thresholds. On 250 safe prompts, refusals rise modestly to 4.0%. Joint optimization improves macro-F1 to 0.9512.
- Input filtering reduces harmful answers to 2.4%
- Tiered filtering yields 5.6% harmful answers and 4.0% benign refusals
- 79 prompts require output review on average
- Macro-F1 improves from 0.9351 to 0.9512
- Modest gains over single-gate in matched budgets
- AI models
- Qwen2.5-1.5B
From the original article
1 min read at arxiv.orgExternal safety filters can prevent unsafe requests from reaching a large language model (LLM), but may also reject legitimate queries. We present an end-to-end empirical evaluation of Jev, a probabilistic decision model, as a safety guardrail for a Qwen2. 5-1.
By Olivia Cognetti · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.