BabyLM Challenge 2024/2025 Dataset

BabyLM Challenge 2024/2025数据集是由IBM研究院、麻省理工学院等机构联合创建，旨在为研究者提供一个样本效率高的预训练环境，模拟人类语言发展过程。数据集包含1亿词，分为文本和图像-文本多模态数据，其中50%为纯文本，50%为图像-文本对。创建过程中，数据来源多样，包括儿童语言数据、对话、儿童故事等，确保了数据的丰富性和多样性。该数据集主要应用于语言模型预训练和认知建模研究，旨在解决在有限数据条件下优化预训练的问题，推动相关领域的技术发展。

The BabyLM Challenge 2024/2025 Dataset was jointly developed by institutions including IBM Research and the Massachusetts Institute of Technology (MIT). Its core objective is to provide researchers with a sample-efficient pre-training environment that simulates the process of human language development. The dataset contains 100 million words, divided into two categories: pure text and image-text multimodal data, with 50% being pure text and the remaining 50% being image-text pairs. During the construction phase, diverse data sources were utilized, including child language corpora, conversational data, children's stories and more, ensuring the richness and diversity of the dataset. This dataset is primarily applied to research on language model pre-training and cognitive modeling, aiming to address the challenge of optimizing pre-training under limited data conditions and advancing technological progress in related fields.