instructor-embedding

Why we included this project

The practical idea behind Instructor is that you tell the model what the text is for. You pair each text with a short instruction, something like 'Represent the Science title' or 'Represent this sentence for duplicate retrieval,' and it returns a vector shaped for that task instead of a generic one. That makes it a good fit when a single model has to handle retrieval, clustering, classification, and text evaluation without any fine-tuning. The encode API is easy to call from existing Python code, and the base, large, and xl checkpoints let you choose where to sit on the speed-versus-quality curve. If you want to go beyond the released models, the training and MTEB evaluation code is there to adapt the approach to your own data.