OpenAlex Topic Classifications for 49M Scientific Datasets (S-Index Challenge)

Topic classifications for ~49 million scientific dataset metadata records from DataCite and ~52K from EMDB (Electron Microscopy Data Bank), mapped to the OpenAlex topic taxonomy (4,516 topics across 4 hierarchical levels: Domain → Field → Subfield → Topic). Each record is semantically matched to the most relevant OpenAlex topic using neural embedding similarity (BAAI/bge-small-en-v1.5), providing full hierarchical classification with interpretable cosine similarity confidence scores. Format NDJSON files (one JSON object per line) with fields: dataset_id (DOI), topic (id, name, score), subfield, field, domain. Performance 300-500 records/second on AMD Threadripper 32-core Zero crashes across 49M records 100% recall via exact FAISS search Produced for the NIH S-Index Challenge Phase 2 (January 2026).