X3Guard-150M is Axiomic Labs' compact prompt guard, fine-tuned from GPT-X3-150M. It reads a user prompt and returns two independent scores: how likely the prompt is to be harmful, and how likely it is to be adversarial (a jailbreak or prompt-injection attempt). At 149M parameters it is both the smallest and best-performing model we evaluated that gives both classifications.
Throughout this model card, all X3Guard figures are at the default "Balanced" setting unless stated otherwise. Each competitor was benchmarked at its own default decision rule/threshold.
X3Guard-150M exhibits good performance for its size. It exhibits similar performance to GLiGuard-300M, and outperforms Horizon-Labs' content-safety-guard-small at the same size. It scores higher than got-oss-safeguard-20b, Llama Guard 3 1B, and Llama Guard 4 12B.
Granite-Guardian-3.1-2B leads at 0.865, about 17x X3Guard's size, and Qwen3Guard-Gen-0.6B (strict) scores 0.843 at 4x X3Guard's size.
| Organization | Model | Parameters | Mean F1 | WildGuardTest F1 | CircleGuardBench F1 |
|---|---|---|---|---|---|
| IBM | Granite Guardian 3.1 2B | 2,534M | 0.865 | 0.794 | 0.935 |
| Qwen | Qwen3Guard-Gen 0.6B (strict) | 596M | 0.843 | 0.885 | 0.801 |
| ToxicityPrompts | PolyGuard Qwen Smol | 494M | 0.834 | 0.873 | 0.796 |
| Fastino | GLiGuard 300M | 208M | 0.827 | 0.862 | 0.793 |
| Axiomic Labs | X3Guard 150M (Balanced) | 149M | 0.818 | 0.874 | 0.761 |
| hbseong | HarmAug Guard | 435M | 0.812 | 0.876 | 0.748 |
| Qwen | Qwen3Guard-Gen 0.6B (loose) | 596M | 0.786 | 0.863 | 0.710 |
| OpenAI | gpt-oss-safeguard-20b† | ~21,000M | 0.762 | 0.826 | 0.699 |
| Horizon-Labs | content-safety-guard-small | 141M | 0.758 | 0.806 | 0.710 |
| Meta | Llama Guard 3 1B | 1,498M | 0.692 | 0.750 | 0.635 |
| Meta | Llama Guard 4 12B | ~12,000M | 0.662 | 0.741 | 0.582 |
| domenicrosati | deberta-v3-xsmall-beavertails-harmful-qa-classifier | 71M | 0.631 | 0.643 | 0.618 |
| enguard | small-guard-32m-en-prompt-safety-binary-nvidia-aegis | 33M | 0.563 | 0.508 | 0.617 |
| DuoGuard | DuoGuard 0.5B | 494M | 0.543 | 0.629 | 0.457 |
| Intel | toxic-prompt-roberta | 125M | 0.511 | 0.510 | 0.512 |
In adversarial detection, X3Guard 150M ranks fifth, behind Sheltron, ProtectAI's two detectors (these were trained on jackhhao, one of the two benchmarks) and DuoGuard 0.5B.
X3Guard scores ahead of gpt-oss-safeguard-20b, both Llama Prompt Guard 2 models, NeuralTrust prompt-guard-oss-small, IBM's Granite Guardian 3.1 2B and others.
| Organization | Model | Parameters | Mean F1 | jackhhao F1 | TrustAIRLab F1 |
|---|---|---|---|---|---|
| Sheltron AI | prompt-guard-68m | 68M | 0.822 | 0.919 | 0.725 |
| ProtectAI | deberta-v3-small-prompt-injection-v2* | 142M | 0.738 | 0.868 | 0.609 |
| ProtectAI | deberta-v3-base-prompt-injection-v2* | 184M | 0.720 | 0.900 | 0.539 |
| DuoGuard | DuoGuard 0.5B | 494M | 0.710 | 0.830 | 0.589 |
| Axiomic Labs | X3Guard 150M (Balanced) | 149M | 0.673 | 0.931 | 0.416 |
| OpenAI | gpt-oss-safeguard-20b† | ~21,000M | 0.663 | 0.877 | 0.448 |
| Qwen | Qwen3Guard-Gen 0.6B (strict) | 596M | 0.648 | 0.944 | 0.353 |
| Fastino | GLiGuard 300M | 208M | 0.617 | 0.800 | 0.435 |
| Meta | Llama Prompt Guard 2 86M | 279M | 0.605 | 0.977 | 0.232 |
| Meta | Llama Prompt Guard 2 22M | 71M | 0.602 | 0.950 | 0.254 |
| NeuralTrust | prompt-guard-oss-small | 141M | 0.565 | 0.831 | 0.299 |
| IBM | Granite Guardian 3.1 2B | 2,534M | 0.561 | 0.920 | 0.202 |
| marklkelly | bert-tiny-injection-detector | 4M | 0.554 | 0.797 | 0.311 |
* Trained on jackhhao, one of the two adversarial benchmarks, according to its model card, so its jackhhao score is not held-out.
Among models that output both classifications, no model is ahead of X3Guard 150M on both. Granite Guardian 3.1 2B and Qwen3Guard-Gen 0.6B (strict) are ahead on harm but behind on adversarial, DuoGuard 0.5B is ahead on adversarial but far behind on harm, and gpt-oss-safeguard-20b is behind on both.
Notes on the comparison
X3Guard 150M ships with three "strictness" settings. Each applies its own threshold to each output, and a prompt is flagged as harmful or adversarial when its score is at or above that threshold. Every figure outside this section uses Balanced.
| Setting | Harm threshold | Adversarial threshold | Good for |
|---|---|---|---|
| Strict | 0.1 | 0.3 | Where false negatives are unacceptable, with the tradeoff being more false positives. Best for moderation tasks where harmful requests must be stopped whenever possible. |
| Balanced (default) | 0.5 | 0.5 | General chatbots and assistants. |
| Permissive | 0.7 | 0.9 | Use cases where false positives must be kept to a minimum, with the tradeoff being more false negatives. Role-play and creative writing apps, developer tools, security research |
Harm head
| Setting | Recall (harmful caught) | Harmful missed | Benign passed | Benign flagged | Precision | F1 |
|---|---|---|---|---|---|---|
| Strict | 82.5% | 17.5% | 84.2% | 15.8% | 85.5% | 0.832 |
| Balanced | 74.7% | 25.3% | 93.7% | 6.3% | 92.7% | 0.818 |
| Permissive | 69.8% | 30.2% | 95.3% | 4.7% | 94.2% | 0.789 |
Adversarial head
| Setting | Recall (attacks caught) | Attacks missed | Benign passed | Benign flagged | Precision | F1 |
|---|---|---|---|---|---|---|
| Strict | 90.0% | 10.0% | 87.0% | 13.0% | 60.3% | 0.661 |
| Balanced | 86.5% | 13.5% | 89.7% | 10.3% | 63.0% | 0.673 |
| Permissive | 78.5% | 21.5% | 93.4% | 6.6% | 66.9% | 0.683 |
These rates are per head. If your application blocks when either head fires, the combined false-flag rate will be higher.
Our headline claims use Balanced (the standard 0.5 threshold). The Strict and Permissive thresholds were chosen with these benchmark results, so treat their numbers as a guide to the trade-off rather than independent claims, and ideally, test the settings on your own traffic before selecting a setting.
| Component | Details |
|---|---|
| Base model | GPT-X3-150M (TX5 hybrid) |
| Parameters | 149M |
| Layers | 24 (18 GDN2 + 6 gated GQA) |
| Hidden size | 576 |
| Tokenizer | LFM2.5 65k |
| Outputs | Two independent sigmoid heads: harm_score, adversarial_score |
| Max prompt length | 1,024 tokens |
| Categories | None; the model does not return harm categories |
Longer prompts keep their first 768 and last 255 tokens, and the middle is dropped. Alternatively, you could use a sliding-window approach to scan the whole prompt. Future Axiomic Guard models will aim to increase this maximum length.
Measured on an NVIDIA RTX 4080:
| Measure | Result |
|---|---|
| Throughput, batch size 32 | 84.5 prompts per second |
| Median latency, one prompt at a time | 33.3 ms (95th percentile: 43.3 ms) |
| Peak GPU memory, one prompt at a time | 1.02 GB |
These figures are for the Transformers (PyTorch) version.
If you are running X3Guard on a CPU, or on individual prompts (rather than batch processing) on a GPU, we recommend the ONNX quants here.
X3Guard 150M was fine-tuned from GPT-X3-150M with a two-output classification head, on a curated mix of real user prompts, public safety and jailbreak datasets, synthetic prompts and AlignmentArena data. Every source was screened against the four benchmarks above before training, and prompts matching or closely resembling a benchmark prompt were removed.
| Dataset | Used for |
|---|---|
| WildGuardMix (train split only) | Harmful, benign and adversarial prompts |
| WildChat-1M and WildChat-4.8M | Real user prompts, including in-the-wild jailbreaks |
| LMSYS-Chat-1M | Real user prompts, including harmful and role-play requests |
| Aegis 2.0 | Harmful and benign prompts with human labels |
| WildJailbreak (harmful subsets) | Harmful requests, plain and wrapped in jailbreaks |
| OR-Bench | Benign prompts that look harmful, to reduce over-flagging |
| HackAPrompt | Prompt-injection attacks |
| Gandalf (ignore instructions, summarization) | Prompt-injection attacks |
| deepset prompt-injections | Prompt-injection attacks and benign prompts |
| Synthetic prompts (Axiomic Labs) | Paired harmful and benign variants, obfuscated requests and hard negatives |
| AlignmentArena (private) | Jailbreak prompts written by players |
The AlignmentArena data is currently private. It contains many highly effective jailbreak prompts, and malicious requests, so it has dual-use potential.
pip install "torch>=2.10" "transformers>=5.5.3,<6" safetensors tokenizers
pip install flash-linear-attention # recommended on GPU; the model falls back to a slower path without it
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from transformers.dynamic_module_utils import get_class_from_dynamic_module
model_id = "AxiomicLabs/X3Guard-150M"
device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained(
model_id,
trust_remote_code=True,
dtype=torch.float32,
).to(device).eval()
classify = get_class_from_dynamic_module("inference.classify", model_id)
prompts = ["What is the capital of France?"]
print(classify(model, tokenizer, prompts)) # Balanced (default)
print(classify(model, tokenizer, prompts, setting="strict")) # or "permissive"
Each prompt returns a dictionary like so:
{"harmful": False, "harm_score": 0.00083, "adversarial": False, "adversarial_score": 0.00004}
classify handles tokenization, the 1,024-token limit and batching (batch_size=16 by default). The thresholds for each setting are in inference.py as SETTINGS; pass threshold= to use a custom threshold for both heads. usage.py in this repository is a runnable version of this example.
The model uses custom code, so trust_remote_code=True is required.
| Benchmark | Task | Source |
|---|---|---|
| WildGuardTest | Harm | Test split of WildGuardMix (AllenAI) |
| CircleGuardBench | Harm | whitecircle/circleguardbench_public (White Circle) |
| jackhhao | Adversarial | jackhhao/jailbreak-classification |
| TrustAIRLab | Adversarial | TrustAIRLab/in-the-wild-jailbreak-prompts (paper code) |
CircleGuardBench results: Dataset provided by WhiteCircle.ai under the White Circle Responsible Use License.
Our thanks to the authors of these benchmarks and of the training datasets above, and to the authors of the compared models.
Many thanks to Joe Le Sage, who provided his hardware for training compute.
@misc{x3guard_2026,
title={X3Guard-150M},
author={Axiomic Labs},
year={2026},
howpublished={\url{https://huggingface.co/AxiomicLabs/X3Guard-150M}},
}
AxiomicLabs/GPT-X3-150M