zhiweizhang988/PubGen
收藏资源简介:
PubGen是一个生成性框架,用于映射生物医学研究领域的科学文献嵌入。该框架包含五种独立的研究任务:1) 使用t-SNE进行降维可视化;2) 基于VAE的生成模型,包括掩码样本重建和Sinkhorn损失;3) 直接从768D嵌入进行多任务分类和分数预测;4) VAE扰动分析,包括稀疏训练、关键词预测和时间趋势检测;5) 使用因果LLM(如Qwen3)进行推理,通过MLP投影器将嵌入向量映射到LLM令牌空间。项目结构详细说明了各任务的代码和数据组织方式。
PubGen is a generative framework for mapping the biomedical research landscape through scientific literature embeddings. It includes five independent research tasks: 1) t-SNE dimensionality reduction for visualization; 2) VAE-based generative modeling with masked sample reconstruction and Sinkhorn loss; 3) direct prediction tasks including multi-task classification and score prediction from original 768D embeddings; 4) VAE perturbation analysis featuring sparse training, keyword prediction, and temporal trend detection; 5) LLM inference using causal models (e.g., Qwen3) with MLP projectors to map embedding vectors to LLM token space. The project structure details the organization of code and data for each task.




