Hierarchical clustering, MeSH-derived labels, Qwen3.5-27B-generated topic names, and atlas-driven research-void rankings for all 28,460,827 PubMed abstracts in the aaekay/pubmed-embeddings release, computed independently for three encoders:

NeuML/pubmedbert-base-embeddings (768-d, biomedical-specialised) Qwen/Qwen3-Embedding-0.6B (1024-d, generalist) BAAI/bge-m3 (1024-d, multilingual generalist)

Joinable to the embedding release by pmid.

Headline numbers… See the full description on the dataset page: https://huggingface.co/datasets/aaekay/pubmed-atlas.