[Feedback Wanted] Low-Latency LLM Guardrails Using a 14M Parameter Discriminator

DeshwalX released electra-small-prompt-injection-v1, a 14M-parameter multi-label classifier for low-latency LLM guardrails, achieving 90% micro F1 on the WildGuard benchmark. The model detects prompt injection, jailbreaks, and harmful content in a single forward pass, offering a faster alternative to large autoregressive detectors.

Hi everyone, I have released DeshwalX/electra-small-prompt-injection-v1 , a multi-label classifier for low-latency LLM guardrails. Many prompt injection detectors rely on large autoregressive models that introduce processing latency. This project tests the baseline efficiency of a compact encoder backbone by fine-tuning google/electra-small-discriminator ~14M parameters to detect four independent safety vectors in a single forward pass. WildGuard Test Benchmark Evaluation: Micro F1: 90.00% | Macro F1: 89.00% Prompt Adversarial Jailbreaks : 1.00 F1 Response Refusal: 0.86 F1 Prompt Harmful: 0.88 F1 Response Harmful: 0.81 F1 Model Repository: DeshwalX/electra-small-prompt-injection-v1 · Hugging Face https://huggingface.co/DeshwalX/electra-small-prompt-injection-v1 I am looking for community feedback to improve Version 1 . Please test the model on your own data and share: Edge cases where the model fails or misses an injection. Examples of false positives on benign inputs. Suggestions for architecture configurations or loss weight balancing for the next iteration.