Lab-Safety-QA

Lab-Safety-QA is a Chinese-language laboratory safety question-answering dataset for training and evaluation in preference learning, safety, and commonsense reasoning.

Explore questions, search keywords, and inspect reference answers in the interactive question explorer.

Dataset Overview

The dataset contains 2,796 questions: 2,098 single-choice questions and 698 multiple-choice questions. Each record includes a question, answer options, reference answer labels, answer text, and source filenames.

The data is provided as UTF-8 JSONL in lab-safety-qa.jsonl, with one JSON object per line.

Data Fields

Field	Type	Description
id	integer	Question identifier.
question	string	Question text in Chinese.
question_type	string	single_choice or multiple_choice.
options	list of objects	Available options, each containing label and text.
answer	list of strings	Reference answer labels; single-choice answers also use a list.
answer_text	list of strings	Option text corresponding to the reference answer labels.
source_files	list of strings	Filenames identifying the records in the original source project.

Usage

Load the dataset from Hugging Face:

from datasets import load_dataset

dataset = load_dataset("choucsan/Lab-Safety-QA", split="train")

Read a local JSONL file with Python:

import json

with open("lab-safety-qa.jsonl", encoding="utf-8") as dataset_file:
    records = [json.loads(line) for line in dataset_file if line.strip()]

Training and Evaluation

• Use the dataset for laboratory safety QA, commonsense understanding, and multiple-choice training and evaluation.
• The QA records can serve as a starting point for preference learning. The dataset does not include chosen / rejected preference pairs; construct and validate these separately before preference training.
• The complete dataset is distributed as one file without a predefined train/test partition. The train label in the loading example is a loading convention, not a held-out evaluation split. Create appropriate partitions and prevent overlap between training and evaluation data.
• This release contains text questions, options, and reference answers. Original question images are not included.

Acknowledgments and Data Source

We thank KY-zhang-X for sharing the original open-source question bank. This dataset is adapted from the project's data directory and organized into a structured JSONL format.

This is an unofficial learning resource. Reference answers are inherited from the original source and have not been professionally validated. They must not replace current laboratory safety procedures or professional guidance, and should not be used directly for safety-critical decisions.

License

The original project is distributed under the MIT License. The original license notice is preserved in LICENSE.