| license: cc-by-4.0 | |
| task_categories: | |
| - image-text-to-text | |
| tags: | |
| - safety | |
| - multimodal | |
| - implicit-risk | |
| # TriggerBench | |
| TriggerBench is the first dataset explicitly modeling risk compositionality for multimodal implicit hazards. It contains 5,600 paired multimodal instances across 9 safety categories, with counterfactual contrastive pairs that decouple Key Elements and Trigger Elements to eliminate risk residues and encourage genuine logical deduction rather than superficial pattern matching. | |
| The dataset is introduced in the paper [ThinkingGuard: Decoding Implicit Hazards via Step-by-Step Risk Attribution in Multimodal Large Language Models](https://huggingface.co/papers/2609.36562). The official code and evaluation scripts are available in the [GitHub repository](https://github.com/FroggyChen/ThinkingGuard). | |
| Dataset details: | |
| - 5,600 paired multimodal instances | |
| - 3,864 train / 1,736 test | |
| - 9 safety categories | |
| - Counterfactual contrastive pairs separating Key Elements and Trigger Elements | |
| Related model weights: [ThinkingGuard-8B](https://huggingface.co/Sandy124/ThinkingGuard-8B) |