3 ms·
how well would this work for 762 dimensional vectors? At 3072, they're starting with such a high number of dimensions that the accuracy loss may not be represen
by crucio 3y ago
how well would this work for 762 dimensional vectors? At 3072, they're starting with such a high number of dimensions that the accuracy loss may not be representative of what others would see
- esafak 3y agoYou'd have to look at the precision-recall curves for your data set and make the trade-off. There are studies on this topic.
- mediaman 3y agoGenerally, it seems that people are starting to see more problems when making vectors of fewer than 1,000 dimensions binary.
- blublubdub 3y agoseems to work pretty well check this out: https://huggingface.co/blog/embedding-quantization https://huggingface.co/blog/embedding-quantization