Instructor

Instruction-finetuned text embeddings (ACL 2023 Findings, legacy anchor): one embedder that produces task- and domain-aware vectors from a natural-language instruction, trained on 330 diverse tasks. At release it set state of the art on 70 embedding benchmarks with an order of magnitude fewer parameters than prior SOTA.

Included as the load-bearing precedent for hkunlp/XLANG's retrieval line: instructor models still see roughly 180K monthly HuggingFace downloads years later, and the instruction-following-embedder idea is now standard across commercial embedding APIs. Built on GTR (T5-based) initializations, not pretrained from scratch.