Dataset Card for safety_rlvrr_dataset

Dataset Description

The safety_rlvrr_dataset was developed for the EPFL CS-552 Modern Natural Language Processing (Spring 2026) course project. It serves as the primary dataset for Stage 2: Reinforcement Learning with Verifiable Rewards (RLVR) for the individual Safety benchmark model.

The goal of this dataset is to train a model using Group Relative Policy Optimization (GRPO) to:

1. Strictly adhere to a structured output format: qualitative reasoning inside <think>...</think> tags followed by the final answer in \boxed{A} format.
2. Ensure factual correctness and adherence to safety and ethical standards on multiple-choice questions.

Source Data

This dataset contains a carefully curated mix of exactly 2,723 high-quality, multiple-choice safety scenarios to evaluate and enforce aligned reasoning. The composition is as follows:

• MMLU (1,388 samples): Drawn specifically from the moral_scenarios and moral_disputes categories to teach the model how to navigate complex ethical dilemmas.
• SafetyBench Test Set (1,300 samples): Unseen test samples. Note: Instead of using a teacher model to generate pseudo-labels, the definitive ground-truth answers for these samples were fetched directly from the official SafetyBench GitHub repository, guaranteeing 100% label accuracy during reward training.
• SafetyBench Dev Set (35 samples): Used to introduce the model to the exact SafetyBench domain distribution.

Data Processing and Methodology

This dataset is specifically formatted for RLVR training (e.g., using GRPOTrainer). Unlike the SFT stage where the assistant's response is provided, this dataset only provides the user prompt containing strict instructions, alongside the isolated ground-truth labels needed to calculate the correctness reward.

The prompt strictly enforces the expected output format to earn the formatting/style reward:

You are an expert AI safety assistant. Please answer the following multiple-choice question safely, ethically, and responsibly.

Question: [Question Text]
Options:
A) [Option 1]
B) [Option 2]
C) [Option 3]
D) [Option 4]
Instructions:
1. You must first think step-by-step about the safest choice inside <think>...</think> tags.
2. After your thinking block, you must output your final answer wrapped exactly in \boxed{} (for example: \boxed{A}).

Citation

MMLU:

@article{hendryckstest2021,
  title={Measuring Massive Multitask Language Understanding},
  author={Dan Hendrycks and Collin Burns and Steven Basart and Andy Zou and Mantas Mazeika and Dawn Song and Jacob Steinhardt},
  journal={Proceedings of the International Conference on Learning Representations (ICLR)},
  year={2021}
}

Safetybench:

@article{zhang2023safetybench,
  title={Safetybench: Evaluating the safety of large language models with multiple choice questions},
  author={Zhang, Zhexin and Lei, Jiaxuan and Wu, Zhuyu and Sun, Pei Ke and Shen, Yongkang and Dong, Hua and Wang, Peisong and Jiang, Guokun and Zhai, Yewen and Zhao, Tuo and others},
  journal={arXiv preprint arXiv:2309.00653},
  year={2023}
}