Effective guardrail models act as a safety layer between the user and the model, preventing nefarious user requests from reaching the model and harmful responses from reaching the user. Current state-of-the-art guardrail models are effective at safety moderation, but they're built on large generative models with billions of parameters. They use text generation to solve what is fundamentally a classification problem, which makes them slow and costly to run at production scale.

In this post we introduce GLiGuard, a 300 million parameter safety moderation model that evaluates multiple safety dimensions in a single pass. Across nine safety benchmarks, GLiGuard’s accuracy matches or exceeds models that are 23 to 90 times its size while running up to 16 times faster.

GLiGuard model weights are available under the Apache 2.0 license on the Hugging Face Hub and the model is available for inference on Pioneer.

Drawbacks of using decoder LLMs for safety moderation

Compounding the problem, most guardrail models need to assess inputs across multiple safety dimensions: what type of harm is present, whether the user prompt is attempting to bypass safety training, whether the model's response is itself unsafe, and so on.Because decoder models generate output sequentially, these assessments are typically produced one after another, and latency compounds as more criteria are evaluated.

GLiGuard performs four safety moderation tasks in a single pass

Instead of generating tokens, GLiGuard encodes both the input text and task definitions (labels) together. These are then fed to the model, which scores every label simultaneously in a single forward pass and returns the highest-scoring label for each task. Because all tasks and their candidate labels are part of the input itself, evaluating additional safety dimensions doesn't add latency; it simply means including more labels in the input.

GLiGuard can be used to filter both user queries and model responses for harmful content and handles four moderation tasks concurrently, all evaluated together in one forward pass:

1. Safety classification (safe / unsafe). Is the text safe or not? Applied to both user prompts before generation and model responses after generation.

2. Jailbreak strategy detection (11 strategies). Is the prompt trying to bypass the model's safety training, and if so, how? Strategies include prompt injection, roleplay bypass, instruction override, social engineering, and others. If any jailbreak strategy is detected, the prompt is automatically flagged as unsafe.

3. Harm category detection (14 categories). What type of harm is present? Categories include violence, sexual content, hate speech, PII exposure, misinformation, child safety, copyright violation, and others. A single input can trigger multiple categories at once.

4. Refusal detection (compliance / refusal). Did the model refuse the user's request or comply with it? This is tracked separately because it helps measure over-refusal (when a model refuses safe requests) and detect false compliance (when a model appears to comply but doesn't). If a refusal is detected, the response is automatically marked as safe.

GLiGuard distinguishes itself from decoder models in that evaluates all four of these tasks in a single forward pass, where decoder-based guard models would need to generate each assessment sequentially. This architectural difference has direct implications for both speed and cost, which we'll look at in the next section alongside GLiGuard's benchmark performance against other state-of-the-art guardrail models.

How we trained GLiGuard

Training data

We trained GLiGuard on a mixture of human annotated and synthetically generated training data. For prompt safety, response safety, and refusal detection, we used WildGuardTrain, a dataset of 87,000 human-annotated examples.

For harm category and jailbreak strategy detection, we generated labels for the unsafe samples using GPT-4.1. During early training, the model struggled to distinguish between similar harm categories like toxic speech and violence, so we used Pioneer to generate supplemental synthetic data with edge cases targeting these fine-grained distinctions. Full training details are available in our paper.

Fine-tuning method

GLiGuard was trained via full fine-tuning of the GLiNER2-base-v1 checkpoint for 20 epochs using the AdamW optimizer. Full hyperparameter details are available in the paper.

How we evaluated GLiGuard

We evaluated GLiGuard on two primary aspects: accuracy and speed. For accuracy, we evaluated GLiGuard across nine established safety benchmarks that cover both prompt and response classification. These benchmarks collectively test whether a model can identify harmful content, withstand adversarial attacks, distinguish between different types of harm, and avoid over-flagging safe content.

We compared GLiGuard against six decoder-based guard models that represent the current state of the art: LlamaGuard4 (12B), WildGuard (7B), ShieldGemma (27B), NemoGuard (8B), PolyGuard (7B), and Qwen3Guard (8B). All results use macro-averaged F1, a standard metric that balances precision and recall, where higher is better and 100 indicates a perfect score.

We also benchmarked latency and throughput on a single NVIDIA A100 GPU against three decoder-based guard models: Qwen3Guard-4B, Qwen3Guard-8B, and ShieldGemma-27B. This measured how many samples each model could process per second and how long a single request takes across a range of batch sizes and sequence lengths.

Accuracy comparable to models 23-90x larger

Across all nine benchmarks, GLiGuard's accuracy matches that of significantly larger decoder models.

• GLiGuard scores 87.7 average F1 on prompt classification, within 1.7 points of the best model (PolyGuard-Qwen at 89.4)

• Second-highest average F1 on response classification (82.7), behind only Qwen3Guard-8B (84.1)

• Outperforms LlamaGuard4-12B, ShieldGemma-27B, and NemoGuard-8B despite being 23-90× smaller

These results suggest that the accuracy gap between small encoder models and large decoder-based guard models is much narrower than their difference in parameter count would imply. GLiGuard's performance indicates that much of the classification ability needed for effective safety moderation can be captured in a far more compact architecture.

Up to 16x faster than current SOTA models

To measure GLiGuard’s efficiency, we compared its throughput and latency to other SOTA guardrail models.

• GLiGuard achieves up to 16.2× higher throughput (133 vs. 8.2 Avg. F1 samples/s at batch size 4)

• GLiGuard achieves up to 16.6× lower latency: (26 ms vs. 426 ms at sequence length 64)

• No other model achieves comparable F1 at a similar parameter count

Overall, these results demonstrate that GLiGuard offers a favorable trade-off between effectiveness and efficiency. Since guardrail models evaluate every user input and every model output, even small latency increases compound quickly as conversations grow. Both GLiGuard’s accuracy and efficiency position it as a competitive alternative to larger, slower decoder-based guardrail models.

Conclusion

With GLiGuard, we demonstrate that effective safety moderation doesn't require models with billions of parameters. And by open sourcing our model, we’re democratizing access to effective and economical safety moderation to anyone who needs it.

Try GLiGuard in Pioneer

GLiGuard is now available for inference on Pioneer.

Additional resources