SRR: Situational Risk Reasoning Dataset

SRR is a supervised fine-tuning dataset for post-training vision-language models (VLMs). It is designed to improve the ability of VLMs to reason about situational risks, where the safety of a user request depends on the interaction between the user's intended activity and the visual scenario.

Unlike conventional safety datasets that mainly focus on whether the user instruction itself is harmful, SRR targets cases where an instruction may be benign in isolation but unsafe under a specific visual context.

Dataset Usage

SRR is intended for supervised fine-tuning (SFT) of VLMs during post-training. The dataset follows a ShareGPT-style conversation format and can be used with common VLM training frameworks such as LLaMA-Factory.

Each example contains:

• image: the visual input;
• messages: a ShareGPT-style list of conversation turns, including the user instruction and the assistant response.

The assistant response is designed to guide the model to reason about the relationship between the activity and the scenario before producing the final answer.

Data Format

Each sample follows the format below:

{
  "image": "<image>",
  "messages": [
    {
      "role": "user",
      "content": "<image> User instruction here."
    },
    {
      "role": "assistant",
      "content": "<think>Safety reasoning here.</think>\n<judgement>Safe or Unsafe</judgement>\n<answer>Final response here.</answer>"
    }
  ]
}