54 1 month ago

Turkish enterprise guardrail classifier — flags jailbreak, cross-tenant data access and content risk across 9 categories. Fine-tuned Qwen3-1.7B (Q8_0). F1 87% (TR). Needs full category definitions + few-shot in the prompt -- >see readme.

c4a021ed15ec · 50B
{
"stop": [
"<|im_end|>"
],
"temperature": 0
}