- Apache 2.0
- English and German
Tiger Guard
Content moderation
Labels text SAFE or UNSAFE: abuse, hate, threats and harmful requests in LLM conversations. For prompts, responses and comments, in English and German.
patronus-studio/tiger-guard-content-moderationProfile.
Task
Content moderation
Abuse, hate and harmful requests in prompts, responses and user text
Signal
UNSAFE
- Parameters
- 308M parameters
- Base
- mmBERT-base
- Evaluation
- Accuracy 88.8% on the 1,500-prompt benchmark (HateCheck, Intuit, Aegis 1.0) with 46 false alarms on 750 harmless prompts; 83.3% of disguised harmful requests caught in the held-out test in English, 75.3% in German.
- Formats
- ONNX FP32 1.23 GB · FP16 615 MB · INT8 308 MB · INT8 with INT4 embeddings 237 MB · x86 builds 510 MB and 242 MB
- Variants and siblings
Tiger Guard Small 141M parameters on mmBERT-small, 106 MB as INT8 with INT4 embeddings. Accuracy 87.6% on the same benchmark with less than half the memory.
- License
- Apache 2.0
- Hugging Face
patronus-studio/tiger-guard-content-moderation
How it works: Tiger Guard.
Input
Comment under the post: Shut your stupid mouth, you worthless idiot, nobody here wants you around.
Output
Application rule: Hold the comment for review
Evaluation.
Model card with protocol88.8%
Accuracy 88.8% on the 1,500-prompt benchmark (HateCheck, Intuit, Aegis 1.0) with 46 false alarms on 750 harmless prompts; 83.3% of disguised harmful requests caught in the held-out test in English, 75.3% in German.
Install.
Patronus Ark loads the model and runs it on the CPU. The quickstart with Transformers or ONNX Runtime is in the model card.
pip install patronus-ark
from patronus_ark import SecurityGateway
scanner = SecurityGateway(
categories=["injection", "dlp", "pii"], max_level="l2",
)
scanner.warmup()
scanner.enqueue("Text to inspect")
while (event := scanner.consume_next_event(timeout=1.0)) is not None:
if event["event_type"] == "result":
r = event["result"]
print(r["category"], r["class_name"], r["confidence"])
else:
break