polymathic-ai/LORE-examples
收藏资源简介:
LORE Examples 是一个小型匹配多模态示例数据集,源自LORE,专为MIMIC模型设计。该数据集提供了足够的跨DNA、RNA和蛋白质模态的示例,使用户能够尝试推理、嵌入和生成任务,而无需自行准备数据。每个示例代表一个单一的生物实体(如转录本和/或其蛋白质),包含多个共观察模态。数据行来自MIMIC训练数据的验证分割(即保留集),因此它们处于分布内,并且长度受模型上下文窗口限制。数据集提供两种配置:raw(原始视图,包含模态名称到原始值的映射,如核苷酸/氨基酸字符串)和tokenized(标记化视图,包含预标记的令牌ID列表),两者覆盖相同的行和模态,但标记化视图跳过了运行重型标记器(如BioBERT文本和ESM3结构)的步骤。数据集支持多种模态,包括核酸轨道(如RNA序列、剪接区域)、蛋白质轨道(如氨基酸序列、3D结构)和文本轨道(如上下文描述、生物医学文献)。数据集用于MIMIC模型的输入、嵌入和生成操作,并附有使用示例代码。
A small set of matched multimodal examples from LORE, for the MIMIC model — enough to try inference, embedding, and generation across DNA, RNA, and protein modalities without wiring up your own data. Each example is a single biological entity (a transcript and/or its protein) with several co-observed modalities. Rows are drawn from the held-out (validation) split of MIMICs training data, so they are in-distribution and length-bounded to the models context window. The dataset offers two configurations: raw (with modality names mapping to raw values like nucleotide/AA strings) and tokenized (with pre-tokenized token-id lists), both covering the same rows and modalities, but the tokenized version skips running heavy tokenizers like BioBERT text and ESM3 structure. It includes various modalities such as nucleic tracks (e.g., RNA sequence, splice regions), protein tracks (e.g., amino acid sequence, 3D structure), and text tracks (e.g., context descriptions, biomedical literature). The dataset is intended for input, embedding, and generation with the MIMIC model, and comes with usage code examples.




