Hugging Face's logo

dcarpintero
/
pangolin-guard-base

Text Classification
Transformers
Safetensors
modernbert
ai-safety
safeguards
guardrails
text-embeddings-inference
Model card Files
xet
 Community
pangolin-guard-base / README.md
dcarpintero's picture
dcarpintero
Update README.md
d139d7b verified
preview code
|
Raw History Blame
3.79 kB
metadata 
library_name: transformers
license: apache-2.0
base_model: answerdotai/ModernBERT-base
tags:
  - generated_from_trainer
metrics:
  - f1
  - accuracy
model-index:
  - name: pangolin-guard-base
    results: []

PangolinGuard-Base

LLM applications face critical security challenges in form of prompt injections and jailbreaks. This can result in models leaking sensitive data or deviating from their intended behavior. Existing safeguard models are not fully open and have limited context windows (e.g., only 512 tokens in LlamaGuard).

PangolinGuard is a ModernBERT (Base), lightweight model that discriminates malicious prompts.

🤗 Tech-Blog | GitHub Repo

Intended uses

  • Adding custom, self-hosted safety checks to AI agents and conversational interfaces
  • Topic and content moderation
  • Mitigating risks when connecting AI pipelines to external services

Evaluation data

Evaluated on unseen data from a subset of specialized benchmarks targeting prompt safety and malicious input detection, while testing over-defense behavior:

  • NotInject: Designed to measure over-defense in prompt guard models by including benign inputs enriched with trigger words common in prompt injection attacks.
  • BIPIA: Evaluates privacy invasion attempts and boundary-pushing queries through indirect prompt injection attacks.
  • Wildguard-Benign: Represents legitimate but potentially ambiguous prompts.
  • PINT: Evaluates particularly nuanced prompt injection, jailbreaks, and benign prompts that could be misidentified as malicious.

image/png

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 5e-05
  • train_batch_size: 64
  • eval_batch_size: 32
  • seed: 42
  • optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
  • lr_scheduler_type: linear
  • num_epochs: 2

Training results

Training Loss Epoch Step Validation Loss F1 Accuracy
0.1622 0.1042 100 0.0755 0.9604 0.9741
0.0694 0.2083 200 0.0525 0.9735 0.9828
0.0552 0.3125 300 0.0857 0.9696 0.9810
0.0535 0.4167 400 0.0345 0.9825 0.9889
0.0371 0.5208 500 0.0343 0.9821 0.9887
0.0402 0.625 600 0.0344 0.9836 0.9894
0.037 0.7292 700 0.0282 0.9869 0.9917
0.0265 0.8333 800 0.0229 0.9895 0.9933
0.0285 0.9375 900 0.0240 0.9885 0.9926
0.0191 1.0417 1000 0.0220 0.9908 0.9941
0.0134 1.1458 1100 0.0228 0.9911 0.9943
0.0124 1.25 1200 0.0230 0.9898 0.9935
0.0136 1.3542 1300 0.0212 0.9910 0.9943
0.0088 1.4583 1400 0.0229 0.9911 0.9943
0.0115 1.5625 1500 0.0211 0.9922 0.9950
0.0058 1.6667 1600 0.0233 0.9920 0.9949
0.0119 1.7708 1700 0.0199 0.9916 0.9946
0.0072 1.875 1800 0.0206 0.9925 0.9952
0.007 1.9792 1900 0.0196 0.9923 0.9950

Framework versions

  • Transformers 4.50.0
  • Pytorch 2.6.0+cu124
  • Datasets 3.4.1
  • Tokenizers 0.21.1