Adapting embedding models to new tasks usually requires training data that captures task-specific information needs and relevance criteria. But manually designing such data is expensive and does not scale well, according to a new paper on arXiv.
The paper introduces SYNTRA, described as a novel method for self-improving data synthesis. The abstract suggests that SYNTRA aims to generate the needed training data automatically, though the full details of the approach are not included in the abstract itself.
If successful, SYNTRA could address a major bottleneck in embedding model deployment, making it easier to adapt models to new domains without extensive manual effort. The authors position the work as a step toward scaling embedding models across a wider range of tasks.