Jinja-2

The 'Jinja-2' dataset/benchmark consists of executable chat templates used to evaluate the effectiveness of inference-time backdoor attacks on language models, specifically assessing their impact on factual accuracy and the ability to emit attacker-controlled outputs.

Papers using Jinja-2 (1)

Inference-Time Backdoors via Chat Templates: From LLM Supply Chains to Agentic System Compromise2026